2 * Low-level SPU handling
4 * (C) Copyright IBM Deutschland Entwicklung GmbH 2005
6 * Author: Arnd Bergmann <arndb@de.ibm.com>
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)
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.
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.
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>
33 #include <linux/mutex.h>
34 #include <linux/linux_logo.h>
36 #include <asm/spu_priv1.h>
37 #include <asm/spu_csa.h>
41 const struct spu_management_ops *spu_management_ops;
42 EXPORT_SYMBOL_GPL(spu_management_ops);
44 const struct spu_priv1_ops *spu_priv1_ops;
45 EXPORT_SYMBOL_GPL(spu_priv1_ops);
47 struct cbe_spu_info cbe_spu_info[MAX_NUMNODES];
48 EXPORT_SYMBOL_GPL(cbe_spu_info);
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
55 EXPORT_SYMBOL_GPL(force_sig_info);
58 * Protects cbe_spu_info and spu->number.
60 static DEFINE_SPINLOCK(spu_lock);
63 * List of all spus in the system.
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.
70 * In addition spu_full_list_lock protects all assignmens to
73 static LIST_HEAD(spu_full_list);
74 static DEFINE_SPINLOCK(spu_full_list_lock);
75 static DEFINE_MUTEX(spu_full_list_mutex);
81 void spu_invalidate_slbs(struct spu *spu)
83 struct spu_priv2 __iomem *priv2 = spu->priv2;
86 spin_lock_irqsave(&spu->register_lock, flags);
87 if (spu_mfc_sr1_get(spu) & MFC_STATE1_RELOCATE_MASK)
88 out_be64(&priv2->slb_invalidate_all_W, 0UL);
89 spin_unlock_irqrestore(&spu->register_lock, flags);
91 EXPORT_SYMBOL_GPL(spu_invalidate_slbs);
93 /* This is called by the MM core when a segment size is changed, to
94 * request a flush of all the SPEs using a given mm
96 void spu_flush_all_slbs(struct mm_struct *mm)
101 spin_lock_irqsave(&spu_full_list_lock, flags);
102 list_for_each_entry(spu, &spu_full_list, full_list) {
104 spu_invalidate_slbs(spu);
106 spin_unlock_irqrestore(&spu_full_list_lock, flags);
109 /* The hack below stinks... try to do something better one of
110 * these days... Does it even work properly with NR_CPUS == 1 ?
112 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
114 int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
116 /* Global TLBIE broadcast required with SPEs. */
117 __cpus_setall(&mm->cpu_vm_mask, nr);
120 void spu_associate_mm(struct spu *spu, struct mm_struct *mm)
124 spin_lock_irqsave(&spu_full_list_lock, flags);
126 spin_unlock_irqrestore(&spu_full_list_lock, flags);
128 mm_needs_global_tlbie(mm);
130 EXPORT_SYMBOL_GPL(spu_associate_mm);
132 int spu_64k_pages_available(void)
134 return mmu_psize_defs[MMU_PAGE_64K].shift != 0;
136 EXPORT_SYMBOL_GPL(spu_64k_pages_available);
138 static void spu_restart_dma(struct spu *spu)
140 struct spu_priv2 __iomem *priv2 = spu->priv2;
142 if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
143 out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
146 static inline void spu_load_slb(struct spu *spu, int slbe, struct spu_slb *slb)
148 struct spu_priv2 __iomem *priv2 = spu->priv2;
150 pr_debug("%s: adding SLB[%d] 0x%016lx 0x%016lx\n",
151 __func__, slbe, slb->vsid, slb->esid);
153 out_be64(&priv2->slb_index_W, slbe);
154 /* set invalid before writing vsid */
155 out_be64(&priv2->slb_esid_RW, 0);
156 /* now it's safe to write the vsid */
157 out_be64(&priv2->slb_vsid_RW, slb->vsid);
158 /* setting the new esid makes the entry valid again */
159 out_be64(&priv2->slb_esid_RW, slb->esid);
162 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
164 struct mm_struct *mm = spu->mm;
168 pr_debug("%s\n", __func__);
170 slb.esid = (ea & ESID_MASK) | SLB_ESID_V;
172 switch(REGION_ID(ea)) {
174 #ifdef CONFIG_PPC_MM_SLICES
175 psize = get_slice_psize(mm, ea);
177 psize = mm->context.user_psize;
179 slb.vsid = (get_vsid(mm->context.id, ea, MMU_SEGSIZE_256M)
180 << SLB_VSID_SHIFT) | SLB_VSID_USER;
182 case VMALLOC_REGION_ID:
183 if (ea < VMALLOC_END)
184 psize = mmu_vmalloc_psize;
186 psize = mmu_io_psize;
187 slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
188 << SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
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;
196 /* Future: support kernel segments so that drivers
199 pr_debug("invalid region access at %016lx\n", ea);
202 slb.vsid |= mmu_psize_defs[psize].sllp;
204 spu_load_slb(spu, spu->slb_replace, &slb);
207 if (spu->slb_replace >= 8)
208 spu->slb_replace = 0;
210 spu_restart_dma(spu);
211 spu->stats.slb_flt++;
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)
218 pr_debug("%s, %lx, %lx\n", __func__, dsisr, ea);
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);
229 spu->class_1_dar = ea;
230 spu->class_1_dsisr = dsisr;
232 spu->stop_callback(spu, 1);
234 spu->class_1_dar = 0;
235 spu->class_1_dsisr = 0;
240 static void __spu_kernel_slb(void *addr, struct spu_slb *slb)
242 unsigned long ea = (unsigned long)addr;
245 if (REGION_ID(ea) == KERNEL_REGION_ID)
246 llp = mmu_psize_defs[mmu_linear_psize].sllp;
248 llp = mmu_psize_defs[mmu_virtual_psize].sllp;
250 slb->vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M) << SLB_VSID_SHIFT) |
251 SLB_VSID_KERNEL | llp;
252 slb->esid = (ea & ESID_MASK) | SLB_ESID_V;
256 * Given an array of @nr_slbs SLB entries, @slbs, return non-zero if the
257 * address @new_addr is present.
259 static inline int __slb_present(struct spu_slb *slbs, int nr_slbs,
262 unsigned long ea = (unsigned long)new_addr;
265 for (i = 0; i < nr_slbs; i++)
266 if (!((slbs[i].esid ^ ea) & ESID_MASK))
273 * Setup the SPU kernel SLBs, in preparation for a context save/restore. We
274 * need to map both the context save area, and the save/restore code.
276 * Because the lscsa and code may cross segment boundaires, we check to see
277 * if mappings are required for the start and end of each range. We currently
278 * assume that the mappings are smaller that one segment - if not, something
279 * is seriously wrong.
281 void spu_setup_kernel_slbs(struct spu *spu, struct spu_lscsa *lscsa,
282 void *code, int code_size)
284 struct spu_slb slbs[4];
286 /* start and end addresses of both mappings */
288 lscsa, (void *)lscsa + sizeof(*lscsa) - 1,
289 code, code + code_size - 1
292 /* check the set of addresses, and create a new entry in the slbs array
293 * if there isn't already a SLB for that address */
294 for (i = 0; i < ARRAY_SIZE(addrs); i++) {
295 if (__slb_present(slbs, nr_slbs, addrs[i]))
298 __spu_kernel_slb(addrs[i], &slbs[nr_slbs]);
302 spin_lock_irq(&spu->register_lock);
303 /* Add the set of SLBs */
304 for (i = 0; i < nr_slbs; i++)
305 spu_load_slb(spu, i, &slbs[i]);
306 spin_unlock_irq(&spu->register_lock);
308 EXPORT_SYMBOL_GPL(spu_setup_kernel_slbs);
311 spu_irq_class_0(int irq, void *data)
314 unsigned long stat, mask;
318 spin_lock(&spu->register_lock);
319 mask = spu_int_mask_get(spu, 0);
320 stat = spu_int_stat_get(spu, 0) & mask;
322 spu->class_0_pending |= stat;
323 spu->class_0_dsisr = spu_mfc_dsisr_get(spu);
324 spu->class_0_dar = spu_mfc_dar_get(spu);
325 spin_unlock(&spu->register_lock);
327 spu->stop_callback(spu, 0);
329 spu->class_0_pending = 0;
330 spu->class_0_dsisr = 0;
331 spu->class_0_dar = 0;
333 spu_int_stat_clear(spu, 0, stat);
339 spu_irq_class_1(int irq, void *data)
342 unsigned long stat, mask, dar, dsisr;
346 /* atomically read & clear class1 status. */
347 spin_lock(&spu->register_lock);
348 mask = spu_int_mask_get(spu, 1);
349 stat = spu_int_stat_get(spu, 1) & mask;
350 dar = spu_mfc_dar_get(spu);
351 dsisr = spu_mfc_dsisr_get(spu);
352 if (stat & CLASS1_STORAGE_FAULT_INTR)
353 spu_mfc_dsisr_set(spu, 0ul);
354 spu_int_stat_clear(spu, 1, stat);
356 if (stat & CLASS1_SEGMENT_FAULT_INTR)
357 __spu_trap_data_seg(spu, dar);
359 spin_unlock(&spu->register_lock);
360 pr_debug("%s: %lx %lx %lx %lx\n", __func__, mask, stat,
363 if (stat & CLASS1_STORAGE_FAULT_INTR)
364 __spu_trap_data_map(spu, dar, dsisr);
366 if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_GET_INTR)
369 if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_PUT_INTR)
372 spu->class_1_dsisr = 0;
373 spu->class_1_dar = 0;
375 return stat ? IRQ_HANDLED : IRQ_NONE;
379 spu_irq_class_2(int irq, void *data)
384 const int mailbox_intrs =
385 CLASS2_MAILBOX_THRESHOLD_INTR | CLASS2_MAILBOX_INTR;
388 spin_lock(&spu->register_lock);
389 stat = spu_int_stat_get(spu, 2);
390 mask = spu_int_mask_get(spu, 2);
391 /* ignore interrupts we're not waiting for */
394 /* mailbox interrupts are level triggered. mask them now before
396 if (stat & mailbox_intrs)
397 spu_int_mask_and(spu, 2, ~(stat & mailbox_intrs));
398 /* acknowledge all interrupts before the callbacks */
399 spu_int_stat_clear(spu, 2, stat);
400 spin_unlock(&spu->register_lock);
402 pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
404 if (stat & CLASS2_MAILBOX_INTR)
405 spu->ibox_callback(spu);
407 if (stat & CLASS2_SPU_STOP_INTR)
408 spu->stop_callback(spu, 2);
410 if (stat & CLASS2_SPU_HALT_INTR)
411 spu->stop_callback(spu, 2);
413 if (stat & CLASS2_SPU_DMA_TAG_GROUP_COMPLETE_INTR)
414 spu->mfc_callback(spu);
416 if (stat & CLASS2_MAILBOX_THRESHOLD_INTR)
417 spu->wbox_callback(spu);
419 spu->stats.class2_intr++;
420 return stat ? IRQ_HANDLED : IRQ_NONE;
423 static int spu_request_irqs(struct spu *spu)
427 if (spu->irqs[0] != NO_IRQ) {
428 snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0",
430 ret = request_irq(spu->irqs[0], spu_irq_class_0,
436 if (spu->irqs[1] != NO_IRQ) {
437 snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1",
439 ret = request_irq(spu->irqs[1], spu_irq_class_1,
445 if (spu->irqs[2] != NO_IRQ) {
446 snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2",
448 ret = request_irq(spu->irqs[2], spu_irq_class_2,
457 if (spu->irqs[1] != NO_IRQ)
458 free_irq(spu->irqs[1], spu);
460 if (spu->irqs[0] != NO_IRQ)
461 free_irq(spu->irqs[0], spu);
466 static void spu_free_irqs(struct spu *spu)
468 if (spu->irqs[0] != NO_IRQ)
469 free_irq(spu->irqs[0], spu);
470 if (spu->irqs[1] != NO_IRQ)
471 free_irq(spu->irqs[1], spu);
472 if (spu->irqs[2] != NO_IRQ)
473 free_irq(spu->irqs[2], spu);
476 void spu_init_channels(struct spu *spu)
478 static const struct {
482 { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
483 { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
485 { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
486 { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
487 { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
489 struct spu_priv2 __iomem *priv2;
494 /* initialize all channel data to zero */
495 for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
498 out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
499 for (count = 0; count < zero_list[i].count; count++)
500 out_be64(&priv2->spu_chnldata_RW, 0);
503 /* initialize channel counts to meaningful values */
504 for (i = 0; i < ARRAY_SIZE(count_list); i++) {
505 out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
506 out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
509 EXPORT_SYMBOL_GPL(spu_init_channels);
511 static int spu_shutdown(struct sys_device *sysdev)
513 struct spu *spu = container_of(sysdev, struct spu, sysdev);
516 spu_destroy_spu(spu);
520 static struct sysdev_class spu_sysdev_class = {
522 .shutdown = spu_shutdown,
525 int spu_add_sysdev_attr(struct sysdev_attribute *attr)
529 mutex_lock(&spu_full_list_mutex);
530 list_for_each_entry(spu, &spu_full_list, full_list)
531 sysdev_create_file(&spu->sysdev, attr);
532 mutex_unlock(&spu_full_list_mutex);
536 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr);
538 int spu_add_sysdev_attr_group(struct attribute_group *attrs)
543 mutex_lock(&spu_full_list_mutex);
544 list_for_each_entry(spu, &spu_full_list, full_list) {
545 rc = sysfs_create_group(&spu->sysdev.kobj, attrs);
547 /* we're in trouble here, but try unwinding anyway */
549 printk(KERN_ERR "%s: can't create sysfs group '%s'\n",
550 __func__, attrs->name);
552 list_for_each_entry_continue_reverse(spu,
553 &spu_full_list, full_list)
554 sysfs_remove_group(&spu->sysdev.kobj, attrs);
559 mutex_unlock(&spu_full_list_mutex);
563 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr_group);
566 void spu_remove_sysdev_attr(struct sysdev_attribute *attr)
570 mutex_lock(&spu_full_list_mutex);
571 list_for_each_entry(spu, &spu_full_list, full_list)
572 sysdev_remove_file(&spu->sysdev, attr);
573 mutex_unlock(&spu_full_list_mutex);
575 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr);
577 void spu_remove_sysdev_attr_group(struct attribute_group *attrs)
581 mutex_lock(&spu_full_list_mutex);
582 list_for_each_entry(spu, &spu_full_list, full_list)
583 sysfs_remove_group(&spu->sysdev.kobj, attrs);
584 mutex_unlock(&spu_full_list_mutex);
586 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr_group);
588 static int spu_create_sysdev(struct spu *spu)
592 spu->sysdev.id = spu->number;
593 spu->sysdev.cls = &spu_sysdev_class;
594 ret = sysdev_register(&spu->sysdev);
596 printk(KERN_ERR "Can't register SPU %d with sysfs\n",
601 sysfs_add_device_to_node(&spu->sysdev, spu->node);
606 static int __init create_spu(void *data)
615 spu = kzalloc(sizeof (*spu), GFP_KERNEL);
619 spu->alloc_state = SPU_FREE;
621 spin_lock_init(&spu->register_lock);
622 spin_lock(&spu_lock);
623 spu->number = number++;
624 spin_unlock(&spu_lock);
626 ret = spu_create_spu(spu, data);
631 spu_mfc_sdr_setup(spu);
632 spu_mfc_sr1_set(spu, 0x33);
633 ret = spu_request_irqs(spu);
637 ret = spu_create_sysdev(spu);
641 mutex_lock(&cbe_spu_info[spu->node].list_mutex);
642 list_add(&spu->cbe_list, &cbe_spu_info[spu->node].spus);
643 cbe_spu_info[spu->node].n_spus++;
644 mutex_unlock(&cbe_spu_info[spu->node].list_mutex);
646 mutex_lock(&spu_full_list_mutex);
647 spin_lock_irqsave(&spu_full_list_lock, flags);
648 list_add(&spu->full_list, &spu_full_list);
649 spin_unlock_irqrestore(&spu_full_list_lock, flags);
650 mutex_unlock(&spu_full_list_mutex);
652 spu->stats.util_state = SPU_UTIL_IDLE_LOADED;
654 spu->stats.tstamp = timespec_to_ns(&ts);
656 INIT_LIST_HEAD(&spu->aff_list);
663 spu_destroy_spu(spu);
670 static const char *spu_state_names[] = {
671 "user", "system", "iowait", "idle"
674 static unsigned long long spu_acct_time(struct spu *spu,
675 enum spu_utilization_state state)
678 unsigned long long time = spu->stats.times[state];
681 * If the spu is idle or the context is stopped, utilization
682 * statistics are not updated. Apply the time delta from the
683 * last recorded state of the spu.
685 if (spu->stats.util_state == state) {
687 time += timespec_to_ns(&ts) - spu->stats.tstamp;
690 return time / NSEC_PER_MSEC;
694 static ssize_t spu_stat_show(struct sys_device *sysdev, char *buf)
696 struct spu *spu = container_of(sysdev, struct spu, sysdev);
698 return sprintf(buf, "%s %llu %llu %llu %llu "
699 "%llu %llu %llu %llu %llu %llu %llu %llu\n",
700 spu_state_names[spu->stats.util_state],
701 spu_acct_time(spu, SPU_UTIL_USER),
702 spu_acct_time(spu, SPU_UTIL_SYSTEM),
703 spu_acct_time(spu, SPU_UTIL_IOWAIT),
704 spu_acct_time(spu, SPU_UTIL_IDLE_LOADED),
705 spu->stats.vol_ctx_switch,
706 spu->stats.invol_ctx_switch,
711 spu->stats.class2_intr,
712 spu->stats.libassist);
715 static SYSDEV_ATTR(stat, 0644, spu_stat_show, NULL);
717 static int __init init_spu_base(void)
721 for (i = 0; i < MAX_NUMNODES; i++) {
722 mutex_init(&cbe_spu_info[i].list_mutex);
723 INIT_LIST_HEAD(&cbe_spu_info[i].spus);
726 if (!spu_management_ops)
729 /* create sysdev class for spus */
730 ret = sysdev_class_register(&spu_sysdev_class);
734 ret = spu_enumerate_spus(create_spu);
737 printk(KERN_WARNING "%s: Error initializing spus\n",
739 goto out_unregister_sysdev_class;
744 * We cannot put the forward declaration in
745 * <linux/linux_logo.h> because of conflicting session type
746 * conflicts for const and __initdata with different compiler
749 extern const struct linux_logo logo_spe_clut224;
751 fb_append_extra_logo(&logo_spe_clut224, ret);
754 mutex_lock(&spu_full_list_mutex);
755 xmon_register_spus(&spu_full_list);
756 crash_register_spus(&spu_full_list);
757 mutex_unlock(&spu_full_list_mutex);
758 spu_add_sysdev_attr(&attr_stat);
764 out_unregister_sysdev_class:
765 sysdev_class_unregister(&spu_sysdev_class);
769 module_init(init_spu_base);
771 MODULE_LICENSE("GPL");
772 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");