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 #include <asm/kexec.h>
41 
42 const struct spu_management_ops *spu_management_ops;
43 EXPORT_SYMBOL_GPL(spu_management_ops);
44 
45 const struct spu_priv1_ops *spu_priv1_ops;
46 EXPORT_SYMBOL_GPL(spu_priv1_ops);
47 
48 struct cbe_spu_info cbe_spu_info[MAX_NUMNODES];
49 EXPORT_SYMBOL_GPL(cbe_spu_info);
50 
51 /*
52  * The spufs fault-handling code needs to call force_sig_info to raise signals
53  * on DMA errors. Export it here to avoid general kernel-wide access to this
54  * function
55  */
56 EXPORT_SYMBOL_GPL(force_sig_info);
57 
58 /*
59  * Protects cbe_spu_info and spu->number.
60  */
61 static DEFINE_SPINLOCK(spu_lock);
62 
63 /*
64  * List of all spus in the system.
65  *
66  * This list is iterated by callers from irq context and callers that
67  * want to sleep.  Thus modifications need to be done with both
68  * spu_full_list_lock and spu_full_list_mutex held, while iterating
69  * through it requires either of these locks.
70  *
71  * In addition spu_full_list_lock protects all assignmens to
72  * spu->mm.
73  */
74 static LIST_HEAD(spu_full_list);
75 static DEFINE_SPINLOCK(spu_full_list_lock);
76 static DEFINE_MUTEX(spu_full_list_mutex);
77 
78 struct spu_slb {
79 	u64 esid, vsid;
80 };
81 
spu_invalidate_slbs(struct spu * spu)82 void spu_invalidate_slbs(struct spu *spu)
83 {
84 	struct spu_priv2 __iomem *priv2 = spu->priv2;
85 	unsigned long flags;
86 
87 	spin_lock_irqsave(&spu->register_lock, flags);
88 	if (spu_mfc_sr1_get(spu) & MFC_STATE1_RELOCATE_MASK)
89 		out_be64(&priv2->slb_invalidate_all_W, 0UL);
90 	spin_unlock_irqrestore(&spu->register_lock, flags);
91 }
92 EXPORT_SYMBOL_GPL(spu_invalidate_slbs);
93 
94 /* This is called by the MM core when a segment size is changed, to
95  * request a flush of all the SPEs using a given mm
96  */
spu_flush_all_slbs(struct mm_struct * mm)97 void spu_flush_all_slbs(struct mm_struct *mm)
98 {
99 	struct spu *spu;
100 	unsigned long flags;
101 
102 	spin_lock_irqsave(&spu_full_list_lock, flags);
103 	list_for_each_entry(spu, &spu_full_list, full_list) {
104 		if (spu->mm == mm)
105 			spu_invalidate_slbs(spu);
106 	}
107 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
108 }
109 
110 /* The hack below stinks... try to do something better one of
111  * these days... Does it even work properly with NR_CPUS == 1 ?
112  */
mm_needs_global_tlbie(struct mm_struct * mm)113 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
114 {
115 	int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
116 
117 	/* Global TLBIE broadcast required with SPEs. */
118 	bitmap_fill(cpumask_bits(mm_cpumask(mm)), nr);
119 }
120 
spu_associate_mm(struct spu * spu,struct mm_struct * mm)121 void spu_associate_mm(struct spu *spu, struct mm_struct *mm)
122 {
123 	unsigned long flags;
124 
125 	spin_lock_irqsave(&spu_full_list_lock, flags);
126 	spu->mm = mm;
127 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
128 	if (mm)
129 		mm_needs_global_tlbie(mm);
130 }
131 EXPORT_SYMBOL_GPL(spu_associate_mm);
132 
spu_64k_pages_available(void)133 int spu_64k_pages_available(void)
134 {
135 	return mmu_psize_defs[MMU_PAGE_64K].shift != 0;
136 }
137 EXPORT_SYMBOL_GPL(spu_64k_pages_available);
138 
spu_restart_dma(struct spu * spu)139 static void spu_restart_dma(struct spu *spu)
140 {
141 	struct spu_priv2 __iomem *priv2 = spu->priv2;
142 
143 	if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
144 		out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
145 	else {
146 		set_bit(SPU_CONTEXT_FAULT_PENDING, &spu->flags);
147 		mb();
148 	}
149 }
150 
spu_load_slb(struct spu * spu,int slbe,struct spu_slb * slb)151 static inline void spu_load_slb(struct spu *spu, int slbe, struct spu_slb *slb)
152 {
153 	struct spu_priv2 __iomem *priv2 = spu->priv2;
154 
155 	pr_debug("%s: adding SLB[%d] 0x%016llx 0x%016llx\n",
156 			__func__, slbe, slb->vsid, slb->esid);
157 
158 	out_be64(&priv2->slb_index_W, slbe);
159 	/* set invalid before writing vsid */
160 	out_be64(&priv2->slb_esid_RW, 0);
161 	/* now it's safe to write the vsid */
162 	out_be64(&priv2->slb_vsid_RW, slb->vsid);
163 	/* setting the new esid makes the entry valid again */
164 	out_be64(&priv2->slb_esid_RW, slb->esid);
165 }
166 
__spu_trap_data_seg(struct spu * spu,unsigned long ea)167 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
168 {
169 	struct mm_struct *mm = spu->mm;
170 	struct spu_slb slb;
171 	int psize;
172 
173 	pr_debug("%s\n", __func__);
174 
175 	slb.esid = (ea & ESID_MASK) | SLB_ESID_V;
176 
177 	switch(REGION_ID(ea)) {
178 	case USER_REGION_ID:
179 #ifdef CONFIG_PPC_MM_SLICES
180 		psize = get_slice_psize(mm, ea);
181 #else
182 		psize = mm->context.user_psize;
183 #endif
184 		slb.vsid = (get_vsid(mm->context.id, ea, MMU_SEGSIZE_256M)
185 				<< SLB_VSID_SHIFT) | SLB_VSID_USER;
186 		break;
187 	case VMALLOC_REGION_ID:
188 		if (ea < VMALLOC_END)
189 			psize = mmu_vmalloc_psize;
190 		else
191 			psize = mmu_io_psize;
192 		slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
193 				<< SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
194 		break;
195 	case KERNEL_REGION_ID:
196 		psize = mmu_linear_psize;
197 		slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
198 				<< SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
199 		break;
200 	default:
201 		/* Future: support kernel segments so that drivers
202 		 * can use SPUs.
203 		 */
204 		pr_debug("invalid region access at %016lx\n", ea);
205 		return 1;
206 	}
207 	slb.vsid |= mmu_psize_defs[psize].sllp;
208 
209 	spu_load_slb(spu, spu->slb_replace, &slb);
210 
211 	spu->slb_replace++;
212 	if (spu->slb_replace >= 8)
213 		spu->slb_replace = 0;
214 
215 	spu_restart_dma(spu);
216 	spu->stats.slb_flt++;
217 	return 0;
218 }
219 
220 extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX
__spu_trap_data_map(struct spu * spu,unsigned long ea,u64 dsisr)221 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
222 {
223 	int ret;
224 
225 	pr_debug("%s, %llx, %lx\n", __func__, dsisr, ea);
226 
227 	/*
228 	 * Handle kernel space hash faults immediately. User hash
229 	 * faults need to be deferred to process context.
230 	 */
231 	if ((dsisr & MFC_DSISR_PTE_NOT_FOUND) &&
232 	    (REGION_ID(ea) != USER_REGION_ID)) {
233 
234 		spin_unlock(&spu->register_lock);
235 		ret = hash_page(ea, _PAGE_PRESENT, 0x300);
236 		spin_lock(&spu->register_lock);
237 
238 		if (!ret) {
239 			spu_restart_dma(spu);
240 			return 0;
241 		}
242 	}
243 
244 	spu->class_1_dar = ea;
245 	spu->class_1_dsisr = dsisr;
246 
247 	spu->stop_callback(spu, 1);
248 
249 	spu->class_1_dar = 0;
250 	spu->class_1_dsisr = 0;
251 
252 	return 0;
253 }
254 
__spu_kernel_slb(void * addr,struct spu_slb * slb)255 static void __spu_kernel_slb(void *addr, struct spu_slb *slb)
256 {
257 	unsigned long ea = (unsigned long)addr;
258 	u64 llp;
259 
260 	if (REGION_ID(ea) == KERNEL_REGION_ID)
261 		llp = mmu_psize_defs[mmu_linear_psize].sllp;
262 	else
263 		llp = mmu_psize_defs[mmu_virtual_psize].sllp;
264 
265 	slb->vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M) << SLB_VSID_SHIFT) |
266 		SLB_VSID_KERNEL | llp;
267 	slb->esid = (ea & ESID_MASK) | SLB_ESID_V;
268 }
269 
270 /**
271  * Given an array of @nr_slbs SLB entries, @slbs, return non-zero if the
272  * address @new_addr is present.
273  */
__slb_present(struct spu_slb * slbs,int nr_slbs,void * new_addr)274 static inline int __slb_present(struct spu_slb *slbs, int nr_slbs,
275 		void *new_addr)
276 {
277 	unsigned long ea = (unsigned long)new_addr;
278 	int i;
279 
280 	for (i = 0; i < nr_slbs; i++)
281 		if (!((slbs[i].esid ^ ea) & ESID_MASK))
282 			return 1;
283 
284 	return 0;
285 }
286 
287 /**
288  * Setup the SPU kernel SLBs, in preparation for a context save/restore. We
289  * need to map both the context save area, and the save/restore code.
290  *
291  * Because the lscsa and code may cross segment boundaires, we check to see
292  * if mappings are required for the start and end of each range. We currently
293  * assume that the mappings are smaller that one segment - if not, something
294  * is seriously wrong.
295  */
spu_setup_kernel_slbs(struct spu * spu,struct spu_lscsa * lscsa,void * code,int code_size)296 void spu_setup_kernel_slbs(struct spu *spu, struct spu_lscsa *lscsa,
297 		void *code, int code_size)
298 {
299 	struct spu_slb slbs[4];
300 	int i, nr_slbs = 0;
301 	/* start and end addresses of both mappings */
302 	void *addrs[] = {
303 		lscsa, (void *)lscsa + sizeof(*lscsa) - 1,
304 		code, code + code_size - 1
305 	};
306 
307 	/* check the set of addresses, and create a new entry in the slbs array
308 	 * if there isn't already a SLB for that address */
309 	for (i = 0; i < ARRAY_SIZE(addrs); i++) {
310 		if (__slb_present(slbs, nr_slbs, addrs[i]))
311 			continue;
312 
313 		__spu_kernel_slb(addrs[i], &slbs[nr_slbs]);
314 		nr_slbs++;
315 	}
316 
317 	spin_lock_irq(&spu->register_lock);
318 	/* Add the set of SLBs */
319 	for (i = 0; i < nr_slbs; i++)
320 		spu_load_slb(spu, i, &slbs[i]);
321 	spin_unlock_irq(&spu->register_lock);
322 }
323 EXPORT_SYMBOL_GPL(spu_setup_kernel_slbs);
324 
325 static irqreturn_t
spu_irq_class_0(int irq,void * data)326 spu_irq_class_0(int irq, void *data)
327 {
328 	struct spu *spu;
329 	unsigned long stat, mask;
330 
331 	spu = data;
332 
333 	spin_lock(&spu->register_lock);
334 	mask = spu_int_mask_get(spu, 0);
335 	stat = spu_int_stat_get(spu, 0) & mask;
336 
337 	spu->class_0_pending |= stat;
338 	spu->class_0_dar = spu_mfc_dar_get(spu);
339 	spu->stop_callback(spu, 0);
340 	spu->class_0_pending = 0;
341 	spu->class_0_dar = 0;
342 
343 	spu_int_stat_clear(spu, 0, stat);
344 	spin_unlock(&spu->register_lock);
345 
346 	return IRQ_HANDLED;
347 }
348 
349 static irqreturn_t
spu_irq_class_1(int irq,void * data)350 spu_irq_class_1(int irq, void *data)
351 {
352 	struct spu *spu;
353 	unsigned long stat, mask, dar, dsisr;
354 
355 	spu = data;
356 
357 	/* atomically read & clear class1 status. */
358 	spin_lock(&spu->register_lock);
359 	mask  = spu_int_mask_get(spu, 1);
360 	stat  = spu_int_stat_get(spu, 1) & mask;
361 	dar   = spu_mfc_dar_get(spu);
362 	dsisr = spu_mfc_dsisr_get(spu);
363 	if (stat & CLASS1_STORAGE_FAULT_INTR)
364 		spu_mfc_dsisr_set(spu, 0ul);
365 	spu_int_stat_clear(spu, 1, stat);
366 
367 	pr_debug("%s: %lx %lx %lx %lx\n", __func__, mask, stat,
368 			dar, dsisr);
369 
370 	if (stat & CLASS1_SEGMENT_FAULT_INTR)
371 		__spu_trap_data_seg(spu, dar);
372 
373 	if (stat & CLASS1_STORAGE_FAULT_INTR)
374 		__spu_trap_data_map(spu, dar, dsisr);
375 
376 	if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_GET_INTR)
377 		;
378 
379 	if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_PUT_INTR)
380 		;
381 
382 	spu->class_1_dsisr = 0;
383 	spu->class_1_dar = 0;
384 
385 	spin_unlock(&spu->register_lock);
386 
387 	return stat ? IRQ_HANDLED : IRQ_NONE;
388 }
389 
390 static irqreturn_t
spu_irq_class_2(int irq,void * data)391 spu_irq_class_2(int irq, void *data)
392 {
393 	struct spu *spu;
394 	unsigned long stat;
395 	unsigned long mask;
396 	const int mailbox_intrs =
397 		CLASS2_MAILBOX_THRESHOLD_INTR | CLASS2_MAILBOX_INTR;
398 
399 	spu = data;
400 	spin_lock(&spu->register_lock);
401 	stat = spu_int_stat_get(spu, 2);
402 	mask = spu_int_mask_get(spu, 2);
403 	/* ignore interrupts we're not waiting for */
404 	stat &= mask;
405 	/* mailbox interrupts are level triggered. mask them now before
406 	 * acknowledging */
407 	if (stat & mailbox_intrs)
408 		spu_int_mask_and(spu, 2, ~(stat & mailbox_intrs));
409 	/* acknowledge all interrupts before the callbacks */
410 	spu_int_stat_clear(spu, 2, stat);
411 
412 	pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
413 
414 	if (stat & CLASS2_MAILBOX_INTR)
415 		spu->ibox_callback(spu);
416 
417 	if (stat & CLASS2_SPU_STOP_INTR)
418 		spu->stop_callback(spu, 2);
419 
420 	if (stat & CLASS2_SPU_HALT_INTR)
421 		spu->stop_callback(spu, 2);
422 
423 	if (stat & CLASS2_SPU_DMA_TAG_GROUP_COMPLETE_INTR)
424 		spu->mfc_callback(spu);
425 
426 	if (stat & CLASS2_MAILBOX_THRESHOLD_INTR)
427 		spu->wbox_callback(spu);
428 
429 	spu->stats.class2_intr++;
430 
431 	spin_unlock(&spu->register_lock);
432 
433 	return stat ? IRQ_HANDLED : IRQ_NONE;
434 }
435 
spu_request_irqs(struct spu * spu)436 static int spu_request_irqs(struct spu *spu)
437 {
438 	int ret = 0;
439 
440 	if (spu->irqs[0] != NO_IRQ) {
441 		snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0",
442 			 spu->number);
443 		ret = request_irq(spu->irqs[0], spu_irq_class_0,
444 				  IRQF_DISABLED,
445 				  spu->irq_c0, spu);
446 		if (ret)
447 			goto bail0;
448 	}
449 	if (spu->irqs[1] != NO_IRQ) {
450 		snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1",
451 			 spu->number);
452 		ret = request_irq(spu->irqs[1], spu_irq_class_1,
453 				  IRQF_DISABLED,
454 				  spu->irq_c1, spu);
455 		if (ret)
456 			goto bail1;
457 	}
458 	if (spu->irqs[2] != NO_IRQ) {
459 		snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2",
460 			 spu->number);
461 		ret = request_irq(spu->irqs[2], spu_irq_class_2,
462 				  IRQF_DISABLED,
463 				  spu->irq_c2, spu);
464 		if (ret)
465 			goto bail2;
466 	}
467 	return 0;
468 
469 bail2:
470 	if (spu->irqs[1] != NO_IRQ)
471 		free_irq(spu->irqs[1], spu);
472 bail1:
473 	if (spu->irqs[0] != NO_IRQ)
474 		free_irq(spu->irqs[0], spu);
475 bail0:
476 	return ret;
477 }
478 
spu_free_irqs(struct spu * spu)479 static void spu_free_irqs(struct spu *spu)
480 {
481 	if (spu->irqs[0] != NO_IRQ)
482 		free_irq(spu->irqs[0], spu);
483 	if (spu->irqs[1] != NO_IRQ)
484 		free_irq(spu->irqs[1], spu);
485 	if (spu->irqs[2] != NO_IRQ)
486 		free_irq(spu->irqs[2], spu);
487 }
488 
spu_init_channels(struct spu * spu)489 void spu_init_channels(struct spu *spu)
490 {
491 	static const struct {
492 		 unsigned channel;
493 		 unsigned count;
494 	} zero_list[] = {
495 		{ 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
496 		{ 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
497 	}, count_list[] = {
498 		{ 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
499 		{ 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
500 		{ 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
501 	};
502 	struct spu_priv2 __iomem *priv2;
503 	int i;
504 
505 	priv2 = spu->priv2;
506 
507 	/* initialize all channel data to zero */
508 	for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
509 		int count;
510 
511 		out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
512 		for (count = 0; count < zero_list[i].count; count++)
513 			out_be64(&priv2->spu_chnldata_RW, 0);
514 	}
515 
516 	/* initialize channel counts to meaningful values */
517 	for (i = 0; i < ARRAY_SIZE(count_list); i++) {
518 		out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
519 		out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
520 	}
521 }
522 EXPORT_SYMBOL_GPL(spu_init_channels);
523 
spu_shutdown(struct sys_device * sysdev)524 static int spu_shutdown(struct sys_device *sysdev)
525 {
526 	struct spu *spu = container_of(sysdev, struct spu, sysdev);
527 
528 	spu_free_irqs(spu);
529 	spu_destroy_spu(spu);
530 	return 0;
531 }
532 
533 static struct sysdev_class spu_sysdev_class = {
534 	.name = "spu",
535 	.shutdown = spu_shutdown,
536 };
537 
spu_add_sysdev_attr(struct sysdev_attribute * attr)538 int spu_add_sysdev_attr(struct sysdev_attribute *attr)
539 {
540 	struct spu *spu;
541 
542 	mutex_lock(&spu_full_list_mutex);
543 	list_for_each_entry(spu, &spu_full_list, full_list)
544 		sysdev_create_file(&spu->sysdev, attr);
545 	mutex_unlock(&spu_full_list_mutex);
546 
547 	return 0;
548 }
549 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr);
550 
spu_add_sysdev_attr_group(struct attribute_group * attrs)551 int spu_add_sysdev_attr_group(struct attribute_group *attrs)
552 {
553 	struct spu *spu;
554 	int rc = 0;
555 
556 	mutex_lock(&spu_full_list_mutex);
557 	list_for_each_entry(spu, &spu_full_list, full_list) {
558 		rc = sysfs_create_group(&spu->sysdev.kobj, attrs);
559 
560 		/* we're in trouble here, but try unwinding anyway */
561 		if (rc) {
562 			printk(KERN_ERR "%s: can't create sysfs group '%s'\n",
563 					__func__, attrs->name);
564 
565 			list_for_each_entry_continue_reverse(spu,
566 					&spu_full_list, full_list)
567 				sysfs_remove_group(&spu->sysdev.kobj, attrs);
568 			break;
569 		}
570 	}
571 
572 	mutex_unlock(&spu_full_list_mutex);
573 
574 	return rc;
575 }
576 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr_group);
577 
578 
spu_remove_sysdev_attr(struct sysdev_attribute * attr)579 void spu_remove_sysdev_attr(struct sysdev_attribute *attr)
580 {
581 	struct spu *spu;
582 
583 	mutex_lock(&spu_full_list_mutex);
584 	list_for_each_entry(spu, &spu_full_list, full_list)
585 		sysdev_remove_file(&spu->sysdev, attr);
586 	mutex_unlock(&spu_full_list_mutex);
587 }
588 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr);
589 
spu_remove_sysdev_attr_group(struct attribute_group * attrs)590 void spu_remove_sysdev_attr_group(struct attribute_group *attrs)
591 {
592 	struct spu *spu;
593 
594 	mutex_lock(&spu_full_list_mutex);
595 	list_for_each_entry(spu, &spu_full_list, full_list)
596 		sysfs_remove_group(&spu->sysdev.kobj, attrs);
597 	mutex_unlock(&spu_full_list_mutex);
598 }
599 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr_group);
600 
spu_create_sysdev(struct spu * spu)601 static int spu_create_sysdev(struct spu *spu)
602 {
603 	int ret;
604 
605 	spu->sysdev.id = spu->number;
606 	spu->sysdev.cls = &spu_sysdev_class;
607 	ret = sysdev_register(&spu->sysdev);
608 	if (ret) {
609 		printk(KERN_ERR "Can't register SPU %d with sysfs\n",
610 				spu->number);
611 		return ret;
612 	}
613 
614 	sysfs_add_device_to_node(&spu->sysdev, spu->node);
615 
616 	return 0;
617 }
618 
create_spu(void * data)619 static int __init create_spu(void *data)
620 {
621 	struct spu *spu;
622 	int ret;
623 	static int number;
624 	unsigned long flags;
625 	struct timespec ts;
626 
627 	ret = -ENOMEM;
628 	spu = kzalloc(sizeof (*spu), GFP_KERNEL);
629 	if (!spu)
630 		goto out;
631 
632 	spu->alloc_state = SPU_FREE;
633 
634 	spin_lock_init(&spu->register_lock);
635 	spin_lock(&spu_lock);
636 	spu->number = number++;
637 	spin_unlock(&spu_lock);
638 
639 	ret = spu_create_spu(spu, data);
640 
641 	if (ret)
642 		goto out_free;
643 
644 	spu_mfc_sdr_setup(spu);
645 	spu_mfc_sr1_set(spu, 0x33);
646 	ret = spu_request_irqs(spu);
647 	if (ret)
648 		goto out_destroy;
649 
650 	ret = spu_create_sysdev(spu);
651 	if (ret)
652 		goto out_free_irqs;
653 
654 	mutex_lock(&cbe_spu_info[spu->node].list_mutex);
655 	list_add(&spu->cbe_list, &cbe_spu_info[spu->node].spus);
656 	cbe_spu_info[spu->node].n_spus++;
657 	mutex_unlock(&cbe_spu_info[spu->node].list_mutex);
658 
659 	mutex_lock(&spu_full_list_mutex);
660 	spin_lock_irqsave(&spu_full_list_lock, flags);
661 	list_add(&spu->full_list, &spu_full_list);
662 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
663 	mutex_unlock(&spu_full_list_mutex);
664 
665 	spu->stats.util_state = SPU_UTIL_IDLE_LOADED;
666 	ktime_get_ts(&ts);
667 	spu->stats.tstamp = timespec_to_ns(&ts);
668 
669 	INIT_LIST_HEAD(&spu->aff_list);
670 
671 	goto out;
672 
673 out_free_irqs:
674 	spu_free_irqs(spu);
675 out_destroy:
676 	spu_destroy_spu(spu);
677 out_free:
678 	kfree(spu);
679 out:
680 	return ret;
681 }
682 
683 static const char *spu_state_names[] = {
684 	"user", "system", "iowait", "idle"
685 };
686 
spu_acct_time(struct spu * spu,enum spu_utilization_state state)687 static unsigned long long spu_acct_time(struct spu *spu,
688 		enum spu_utilization_state state)
689 {
690 	struct timespec ts;
691 	unsigned long long time = spu->stats.times[state];
692 
693 	/*
694 	 * If the spu is idle or the context is stopped, utilization
695 	 * statistics are not updated.  Apply the time delta from the
696 	 * last recorded state of the spu.
697 	 */
698 	if (spu->stats.util_state == state) {
699 		ktime_get_ts(&ts);
700 		time += timespec_to_ns(&ts) - spu->stats.tstamp;
701 	}
702 
703 	return time / NSEC_PER_MSEC;
704 }
705 
706 
spu_stat_show(struct sys_device * sysdev,struct sysdev_attribute * attr,char * buf)707 static ssize_t spu_stat_show(struct sys_device *sysdev,
708 				struct sysdev_attribute *attr, char *buf)
709 {
710 	struct spu *spu = container_of(sysdev, struct spu, sysdev);
711 
712 	return sprintf(buf, "%s %llu %llu %llu %llu "
713 		      "%llu %llu %llu %llu %llu %llu %llu %llu\n",
714 		spu_state_names[spu->stats.util_state],
715 		spu_acct_time(spu, SPU_UTIL_USER),
716 		spu_acct_time(spu, SPU_UTIL_SYSTEM),
717 		spu_acct_time(spu, SPU_UTIL_IOWAIT),
718 		spu_acct_time(spu, SPU_UTIL_IDLE_LOADED),
719 		spu->stats.vol_ctx_switch,
720 		spu->stats.invol_ctx_switch,
721 		spu->stats.slb_flt,
722 		spu->stats.hash_flt,
723 		spu->stats.min_flt,
724 		spu->stats.maj_flt,
725 		spu->stats.class2_intr,
726 		spu->stats.libassist);
727 }
728 
729 static SYSDEV_ATTR(stat, 0644, spu_stat_show, NULL);
730 
731 #ifdef CONFIG_KEXEC
732 
733 struct crash_spu_info {
734 	struct spu *spu;
735 	u32 saved_spu_runcntl_RW;
736 	u32 saved_spu_status_R;
737 	u32 saved_spu_npc_RW;
738 	u64 saved_mfc_sr1_RW;
739 	u64 saved_mfc_dar;
740 	u64 saved_mfc_dsisr;
741 };
742 
743 #define CRASH_NUM_SPUS	16	/* Enough for current hardware */
744 static struct crash_spu_info crash_spu_info[CRASH_NUM_SPUS];
745 
crash_kexec_stop_spus(void)746 static void crash_kexec_stop_spus(void)
747 {
748 	struct spu *spu;
749 	int i;
750 	u64 tmp;
751 
752 	for (i = 0; i < CRASH_NUM_SPUS; i++) {
753 		if (!crash_spu_info[i].spu)
754 			continue;
755 
756 		spu = crash_spu_info[i].spu;
757 
758 		crash_spu_info[i].saved_spu_runcntl_RW =
759 			in_be32(&spu->problem->spu_runcntl_RW);
760 		crash_spu_info[i].saved_spu_status_R =
761 			in_be32(&spu->problem->spu_status_R);
762 		crash_spu_info[i].saved_spu_npc_RW =
763 			in_be32(&spu->problem->spu_npc_RW);
764 
765 		crash_spu_info[i].saved_mfc_dar    = spu_mfc_dar_get(spu);
766 		crash_spu_info[i].saved_mfc_dsisr  = spu_mfc_dsisr_get(spu);
767 		tmp = spu_mfc_sr1_get(spu);
768 		crash_spu_info[i].saved_mfc_sr1_RW = tmp;
769 
770 		tmp &= ~MFC_STATE1_MASTER_RUN_CONTROL_MASK;
771 		spu_mfc_sr1_set(spu, tmp);
772 
773 		__delay(200);
774 	}
775 }
776 
crash_register_spus(struct list_head * list)777 static void crash_register_spus(struct list_head *list)
778 {
779 	struct spu *spu;
780 	int ret;
781 
782 	list_for_each_entry(spu, list, full_list) {
783 		if (WARN_ON(spu->number >= CRASH_NUM_SPUS))
784 			continue;
785 
786 		crash_spu_info[spu->number].spu = spu;
787 	}
788 
789 	ret = crash_shutdown_register(&crash_kexec_stop_spus);
790 	if (ret)
791 		printk(KERN_ERR "Could not register SPU crash handler");
792 }
793 
794 #else
crash_register_spus(struct list_head * list)795 static inline void crash_register_spus(struct list_head *list)
796 {
797 }
798 #endif
799 
init_spu_base(void)800 static int __init init_spu_base(void)
801 {
802 	int i, ret = 0;
803 
804 	for (i = 0; i < MAX_NUMNODES; i++) {
805 		mutex_init(&cbe_spu_info[i].list_mutex);
806 		INIT_LIST_HEAD(&cbe_spu_info[i].spus);
807 	}
808 
809 	if (!spu_management_ops)
810 		goto out;
811 
812 	/* create sysdev class for spus */
813 	ret = sysdev_class_register(&spu_sysdev_class);
814 	if (ret)
815 		goto out;
816 
817 	ret = spu_enumerate_spus(create_spu);
818 
819 	if (ret < 0) {
820 		printk(KERN_WARNING "%s: Error initializing spus\n",
821 			__func__);
822 		goto out_unregister_sysdev_class;
823 	}
824 
825 	if (ret > 0)
826 		fb_append_extra_logo(&logo_spe_clut224, ret);
827 
828 	mutex_lock(&spu_full_list_mutex);
829 	xmon_register_spus(&spu_full_list);
830 	crash_register_spus(&spu_full_list);
831 	mutex_unlock(&spu_full_list_mutex);
832 	spu_add_sysdev_attr(&attr_stat);
833 
834 	spu_init_affinity();
835 
836 	return 0;
837 
838  out_unregister_sysdev_class:
839 	sysdev_class_unregister(&spu_sysdev_class);
840  out:
841 	return ret;
842 }
843 module_init(init_spu_base);
844 
845 MODULE_LICENSE("GPL");
846 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");
847