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/poll.h>
29 #include <linux/ptrace.h>
30 #include <linux/slab.h>
31 #include <linux/wait.h>
35 #include <linux/mutex.h>
37 #include <asm/spu_priv1.h>
38 #include <asm/mmu_context.h>
40 #include "interrupt.h"
42 const struct spu_priv1_ops *spu_priv1_ops;
44 EXPORT_SYMBOL_GPL(spu_priv1_ops);
46 static int __spu_trap_invalid_dma(struct spu *spu)
48 pr_debug("%s\n", __FUNCTION__);
49 force_sig(SIGBUS, /* info, */ current);
53 static int __spu_trap_dma_align(struct spu *spu)
55 pr_debug("%s\n", __FUNCTION__);
56 force_sig(SIGBUS, /* info, */ current);
60 static int __spu_trap_error(struct spu *spu)
62 pr_debug("%s\n", __FUNCTION__);
63 force_sig(SIGILL, /* info, */ current);
67 static void spu_restart_dma(struct spu *spu)
69 struct spu_priv2 __iomem *priv2 = spu->priv2;
71 if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
72 out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
75 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
77 struct spu_priv2 __iomem *priv2 = spu->priv2;
78 struct mm_struct *mm = spu->mm;
81 pr_debug("%s\n", __FUNCTION__);
83 if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
84 /* SLBs are pre-loaded for context switch, so
85 * we should never get here!
87 printk("%s: invalid access during switch!\n", __func__);
90 if (!mm || (REGION_ID(ea) != USER_REGION_ID)) {
91 /* Future: support kernel segments so that drivers
94 pr_debug("invalid region access at %016lx\n", ea);
98 esid = (ea & ESID_MASK) | SLB_ESID_V;
99 #ifdef CONFIG_HUGETLB_PAGE
100 if (in_hugepage_area(mm->context, ea))
101 llp = mmu_psize_defs[mmu_huge_psize].sllp;
104 llp = mmu_psize_defs[mmu_virtual_psize].sllp;
105 vsid = (get_vsid(mm->context.id, ea) << SLB_VSID_SHIFT) |
108 out_be64(&priv2->slb_index_W, spu->slb_replace);
109 out_be64(&priv2->slb_vsid_RW, vsid);
110 out_be64(&priv2->slb_esid_RW, esid);
113 if (spu->slb_replace >= 8)
114 spu->slb_replace = 0;
116 spu_restart_dma(spu);
121 extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX
122 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
124 pr_debug("%s, %lx, %lx\n", __FUNCTION__, dsisr, ea);
126 /* Handle kernel space hash faults immediately.
127 User hash faults need to be deferred to process context. */
128 if ((dsisr & MFC_DSISR_PTE_NOT_FOUND)
129 && REGION_ID(ea) != USER_REGION_ID
130 && hash_page(ea, _PAGE_PRESENT, 0x300) == 0) {
131 spu_restart_dma(spu);
135 if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
136 printk("%s: invalid access during switch!\n", __func__);
143 spu->stop_callback(spu);
148 spu_irq_class_0(int irq, void *data, struct pt_regs *regs)
153 spu->class_0_pending = 1;
154 spu->stop_callback(spu);
160 spu_irq_class_0_bottom(struct spu *spu)
162 unsigned long stat, mask;
164 spu->class_0_pending = 0;
166 mask = spu_int_mask_get(spu, 0);
167 stat = spu_int_stat_get(spu, 0);
171 if (stat & 1) /* invalid DMA alignment */
172 __spu_trap_dma_align(spu);
174 if (stat & 2) /* invalid MFC DMA */
175 __spu_trap_invalid_dma(spu);
177 if (stat & 4) /* error on SPU */
178 __spu_trap_error(spu);
180 spu_int_stat_clear(spu, 0, stat);
182 return (stat & 0x7) ? -EIO : 0;
184 EXPORT_SYMBOL_GPL(spu_irq_class_0_bottom);
187 spu_irq_class_1(int irq, void *data, struct pt_regs *regs)
190 unsigned long stat, mask, dar, dsisr;
194 /* atomically read & clear class1 status. */
195 spin_lock(&spu->register_lock);
196 mask = spu_int_mask_get(spu, 1);
197 stat = spu_int_stat_get(spu, 1) & mask;
198 dar = spu_mfc_dar_get(spu);
199 dsisr = spu_mfc_dsisr_get(spu);
200 if (stat & 2) /* mapping fault */
201 spu_mfc_dsisr_set(spu, 0ul);
202 spu_int_stat_clear(spu, 1, stat);
203 spin_unlock(&spu->register_lock);
204 pr_debug("%s: %lx %lx %lx %lx\n", __FUNCTION__, mask, stat,
207 if (stat & 1) /* segment fault */
208 __spu_trap_data_seg(spu, dar);
210 if (stat & 2) { /* mapping fault */
211 __spu_trap_data_map(spu, dar, dsisr);
214 if (stat & 4) /* ls compare & suspend on get */
217 if (stat & 8) /* ls compare & suspend on put */
220 return stat ? IRQ_HANDLED : IRQ_NONE;
222 EXPORT_SYMBOL_GPL(spu_irq_class_1_bottom);
225 spu_irq_class_2(int irq, void *data, struct pt_regs *regs)
232 spin_lock(&spu->register_lock);
233 stat = spu_int_stat_get(spu, 2);
234 mask = spu_int_mask_get(spu, 2);
235 /* ignore interrupts we're not waiting for */
238 * mailbox interrupts (0x1 and 0x10) are level triggered.
239 * mask them now before acknowledging.
242 spu_int_mask_and(spu, 2, ~(stat & 0x11));
243 /* acknowledge all interrupts before the callbacks */
244 spu_int_stat_clear(spu, 2, stat);
245 spin_unlock(&spu->register_lock);
247 pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
249 if (stat & 1) /* PPC core mailbox */
250 spu->ibox_callback(spu);
252 if (stat & 2) /* SPU stop-and-signal */
253 spu->stop_callback(spu);
255 if (stat & 4) /* SPU halted */
256 spu->stop_callback(spu);
258 if (stat & 8) /* DMA tag group complete */
259 spu->mfc_callback(spu);
261 if (stat & 0x10) /* SPU mailbox threshold */
262 spu->wbox_callback(spu);
264 return stat ? IRQ_HANDLED : IRQ_NONE;
268 spu_request_irqs(struct spu *spu)
273 irq_base = IIC_NODE_STRIDE * spu->node + IIC_SPE_OFFSET;
275 snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0", spu->number);
276 ret = request_irq(irq_base + spu->isrc,
277 spu_irq_class_0, SA_INTERRUPT, spu->irq_c0, spu);
281 snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1", spu->number);
282 ret = request_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc,
283 spu_irq_class_1, SA_INTERRUPT, spu->irq_c1, spu);
287 snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2", spu->number);
288 ret = request_irq(irq_base + 2*IIC_CLASS_STRIDE + spu->isrc,
289 spu_irq_class_2, SA_INTERRUPT, spu->irq_c2, spu);
295 free_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc, spu);
297 free_irq(irq_base + spu->isrc, spu);
303 spu_free_irqs(struct spu *spu)
307 irq_base = IIC_NODE_STRIDE * spu->node + IIC_SPE_OFFSET;
309 free_irq(irq_base + spu->isrc, spu);
310 free_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc, spu);
311 free_irq(irq_base + 2*IIC_CLASS_STRIDE + spu->isrc, spu);
314 static LIST_HEAD(spu_list);
315 static DEFINE_MUTEX(spu_mutex);
317 static void spu_init_channels(struct spu *spu)
319 static const struct {
323 { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
324 { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
326 { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
327 { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
328 { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
330 struct spu_priv2 __iomem *priv2;
335 /* initialize all channel data to zero */
336 for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
339 out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
340 for (count = 0; count < zero_list[i].count; count++)
341 out_be64(&priv2->spu_chnldata_RW, 0);
344 /* initialize channel counts to meaningful values */
345 for (i = 0; i < ARRAY_SIZE(count_list); i++) {
346 out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
347 out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
351 struct spu *spu_alloc(void)
355 mutex_lock(&spu_mutex);
356 if (!list_empty(&spu_list)) {
357 spu = list_entry(spu_list.next, struct spu, list);
358 list_del_init(&spu->list);
359 pr_debug("Got SPU %x %d\n", spu->isrc, spu->number);
361 pr_debug("No SPU left\n");
364 mutex_unlock(&spu_mutex);
367 spu_init_channels(spu);
371 EXPORT_SYMBOL_GPL(spu_alloc);
373 void spu_free(struct spu *spu)
375 mutex_lock(&spu_mutex);
376 list_add_tail(&spu->list, &spu_list);
377 mutex_unlock(&spu_mutex);
379 EXPORT_SYMBOL_GPL(spu_free);
381 static int spu_handle_mm_fault(struct spu *spu)
383 struct mm_struct *mm = spu->mm;
384 struct vm_area_struct *vma;
385 u64 ea, dsisr, is_write;
391 if (!IS_VALID_EA(ea)) {
398 if (mm->pgd == NULL) {
402 down_read(&mm->mmap_sem);
403 vma = find_vma(mm, ea);
406 if (vma->vm_start <= ea)
408 if (!(vma->vm_flags & VM_GROWSDOWN))
411 if (expand_stack(vma, ea))
415 is_write = dsisr & MFC_DSISR_ACCESS_PUT;
417 if (!(vma->vm_flags & VM_WRITE))
420 if (dsisr & MFC_DSISR_ACCESS_DENIED)
422 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
426 switch (handle_mm_fault(mm, vma, ea, is_write)) {
433 case VM_FAULT_SIGBUS:
442 up_read(&mm->mmap_sem);
446 up_read(&mm->mmap_sem);
450 int spu_irq_class_1_bottom(struct spu *spu)
452 u64 ea, dsisr, access, error = 0UL;
457 if (dsisr & (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED)) {
460 access = (_PAGE_PRESENT | _PAGE_USER);
461 access |= (dsisr & MFC_DSISR_ACCESS_PUT) ? _PAGE_RW : 0UL;
462 local_irq_save(flags);
463 if (hash_page(ea, access, 0x300) != 0)
464 error |= CLASS1_ENABLE_STORAGE_FAULT_INTR;
465 local_irq_restore(flags);
467 if (error & CLASS1_ENABLE_STORAGE_FAULT_INTR) {
468 if ((ret = spu_handle_mm_fault(spu)) != 0)
469 error |= CLASS1_ENABLE_STORAGE_FAULT_INTR;
471 error &= ~CLASS1_ENABLE_STORAGE_FAULT_INTR;
476 spu_restart_dma(spu);
478 __spu_trap_invalid_dma(spu);
483 static int __init find_spu_node_id(struct device_node *spe)
486 struct device_node *cpu;
487 cpu = spe->parent->parent;
488 id = (unsigned int *)get_property(cpu, "node-id", NULL);
492 static int __init cell_spuprop_present(struct spu *spu, struct device_node *spe,
495 static DEFINE_MUTEX(add_spumem_mutex);
497 struct address_prop {
498 unsigned long address;
500 } __attribute__((packed)) *p;
503 unsigned long start_pfn, nr_pages;
504 struct pglist_data *pgdata;
508 p = (void*)get_property(spe, prop, &proplen);
509 WARN_ON(proplen != sizeof (*p));
511 start_pfn = p->address >> PAGE_SHIFT;
512 nr_pages = ((unsigned long)p->len + PAGE_SIZE - 1) >> PAGE_SHIFT;
514 pgdata = NODE_DATA(spu->nid);
515 zone = pgdata->node_zones;
517 /* XXX rethink locking here */
518 mutex_lock(&add_spumem_mutex);
519 ret = __add_pages(zone, start_pfn, nr_pages);
520 mutex_unlock(&add_spumem_mutex);
525 static void __iomem * __init map_spe_prop(struct spu *spu,
526 struct device_node *n, const char *name)
528 struct address_prop {
529 unsigned long address;
531 } __attribute__((packed)) *prop;
538 p = get_property(n, name, &proplen);
539 if (proplen != sizeof (struct address_prop))
544 err = cell_spuprop_present(spu, n, name);
545 if (err && (err != -EEXIST))
548 ret = ioremap(prop->address, prop->len);
554 static void spu_unmap(struct spu *spu)
558 iounmap(spu->problem);
559 iounmap((u8 __iomem *)spu->local_store);
562 static int __init spu_map_device(struct spu *spu, struct device_node *node)
568 prop = get_property(node, "isrc", NULL);
571 spu->isrc = *(unsigned int *)prop;
573 spu->name = get_property(node, "name", NULL);
577 prop = get_property(node, "local-store", NULL);
580 spu->local_store_phys = *(unsigned long *)prop;
582 /* we use local store as ram, not io memory */
583 spu->local_store = (void __force *)
584 map_spe_prop(spu, node, "local-store");
585 if (!spu->local_store)
588 prop = get_property(node, "problem", NULL);
591 spu->problem_phys = *(unsigned long *)prop;
593 spu->problem= map_spe_prop(spu, node, "problem");
597 spu->priv1= map_spe_prop(spu, node, "priv1");
598 /* priv1 is not available on a hypervisor */
600 spu->priv2= map_spe_prop(spu, node, "priv2");
612 struct sysdev_class spu_sysdev_class = {
616 static ssize_t spu_show_isrc(struct sys_device *sysdev, char *buf)
618 struct spu *spu = container_of(sysdev, struct spu, sysdev);
619 return sprintf(buf, "%d\n", spu->isrc);
622 static SYSDEV_ATTR(isrc, 0400, spu_show_isrc, NULL);
624 extern int attach_sysdev_to_node(struct sys_device *dev, int nid);
626 static int spu_create_sysdev(struct spu *spu)
630 spu->sysdev.id = spu->number;
631 spu->sysdev.cls = &spu_sysdev_class;
632 ret = sysdev_register(&spu->sysdev);
634 printk(KERN_ERR "Can't register SPU %d with sysfs\n",
639 sysdev_create_file(&spu->sysdev, &attr_isrc);
640 sysfs_add_device_to_node(&spu->sysdev, spu->nid);
645 static void spu_destroy_sysdev(struct spu *spu)
647 sysdev_remove_file(&spu->sysdev, &attr_isrc);
648 sysfs_remove_device_from_node(&spu->sysdev, spu->nid);
649 sysdev_unregister(&spu->sysdev);
652 static int __init create_spu(struct device_node *spe)
659 spu = kzalloc(sizeof (*spu), GFP_KERNEL);
663 ret = spu_map_device(spu, spe);
667 spu->node = find_spu_node_id(spe);
668 spu->nid = of_node_to_nid(spe);
671 spin_lock_init(&spu->register_lock);
672 spu_mfc_sdr_set(spu, mfspr(SPRN_SDR1));
673 spu_mfc_sr1_set(spu, 0x33);
674 mutex_lock(&spu_mutex);
676 spu->number = number++;
677 ret = spu_request_irqs(spu);
681 ret = spu_create_sysdev(spu);
685 list_add(&spu->list, &spu_list);
686 mutex_unlock(&spu_mutex);
688 pr_debug(KERN_DEBUG "Using SPE %s %02x %p %p %p %p %d\n",
689 spu->name, spu->isrc, spu->local_store,
690 spu->problem, spu->priv1, spu->priv2, spu->number);
697 mutex_unlock(&spu_mutex);
705 static void destroy_spu(struct spu *spu)
707 list_del_init(&spu->list);
709 spu_destroy_sysdev(spu);
715 static void cleanup_spu_base(void)
717 struct spu *spu, *tmp;
718 mutex_lock(&spu_mutex);
719 list_for_each_entry_safe(spu, tmp, &spu_list, list)
721 mutex_unlock(&spu_mutex);
722 sysdev_class_unregister(&spu_sysdev_class);
724 module_exit(cleanup_spu_base);
726 static int __init init_spu_base(void)
728 struct device_node *node;
731 /* create sysdev class for spus */
732 ret = sysdev_class_register(&spu_sysdev_class);
737 for (node = of_find_node_by_type(NULL, "spe");
738 node; node = of_find_node_by_type(node, "spe")) {
739 ret = create_spu(node);
741 printk(KERN_WARNING "%s: Error initializing %s\n",
742 __FUNCTION__, node->name);
747 /* in some old firmware versions, the spe is called 'spc', so we
748 look for that as well */
749 for (node = of_find_node_by_type(NULL, "spc");
750 node; node = of_find_node_by_type(node, "spc")) {
751 ret = create_spu(node);
753 printk(KERN_WARNING "%s: Error initializing %s\n",
754 __FUNCTION__, node->name);
761 module_init(init_spu_base);
763 MODULE_LICENSE("GPL");
764 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");