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/mmu_context.h>
39 #include "interrupt.h"
41 static int __spu_trap_invalid_dma(struct spu *spu)
43 pr_debug("%s\n", __FUNCTION__);
44 force_sig(SIGBUS, /* info, */ current);
48 static int __spu_trap_dma_align(struct spu *spu)
50 pr_debug("%s\n", __FUNCTION__);
51 force_sig(SIGBUS, /* info, */ current);
55 static int __spu_trap_error(struct spu *spu)
57 pr_debug("%s\n", __FUNCTION__);
58 force_sig(SIGILL, /* info, */ current);
62 static void spu_restart_dma(struct spu *spu)
64 struct spu_priv2 __iomem *priv2 = spu->priv2;
66 if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
67 out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
70 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
72 struct spu_priv2 __iomem *priv2 = spu->priv2;
73 struct mm_struct *mm = spu->mm;
76 pr_debug("%s\n", __FUNCTION__);
78 if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
79 /* SLBs are pre-loaded for context switch, so
80 * we should never get here!
82 printk("%s: invalid access during switch!\n", __func__);
85 if (!mm || (REGION_ID(ea) != USER_REGION_ID)) {
86 /* Future: support kernel segments so that drivers
89 pr_debug("invalid region access at %016lx\n", ea);
93 esid = (ea & ESID_MASK) | SLB_ESID_V;
94 vsid = (get_vsid(mm->context.id, ea) << SLB_VSID_SHIFT) | SLB_VSID_USER;
95 if (in_hugepage_area(mm->context, ea))
98 out_be64(&priv2->slb_index_W, spu->slb_replace);
99 out_be64(&priv2->slb_vsid_RW, vsid);
100 out_be64(&priv2->slb_esid_RW, esid);
103 if (spu->slb_replace >= 8)
104 spu->slb_replace = 0;
106 spu_restart_dma(spu);
111 extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX
112 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
114 pr_debug("%s, %lx, %lx\n", __FUNCTION__, dsisr, ea);
116 /* Handle kernel space hash faults immediately.
117 User hash faults need to be deferred to process context. */
118 if ((dsisr & MFC_DSISR_PTE_NOT_FOUND)
119 && REGION_ID(ea) != USER_REGION_ID
120 && hash_page(ea, _PAGE_PRESENT, 0x300) == 0) {
121 spu_restart_dma(spu);
125 if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
126 printk("%s: invalid access during switch!\n", __func__);
133 if (spu->stop_callback)
134 spu->stop_callback(spu);
138 static int __spu_trap_mailbox(struct spu *spu)
140 if (spu->ibox_callback)
141 spu->ibox_callback(spu);
143 /* atomically disable SPU mailbox interrupts */
144 spin_lock(&spu->register_lock);
145 spu_int_mask_and(spu, 2, ~0x1);
146 spin_unlock(&spu->register_lock);
150 static int __spu_trap_stop(struct spu *spu)
152 pr_debug("%s\n", __FUNCTION__);
153 spu->stop_code = in_be32(&spu->problem->spu_status_R);
154 if (spu->stop_callback)
155 spu->stop_callback(spu);
159 static int __spu_trap_halt(struct spu *spu)
161 pr_debug("%s\n", __FUNCTION__);
162 spu->stop_code = in_be32(&spu->problem->spu_status_R);
163 if (spu->stop_callback)
164 spu->stop_callback(spu);
168 static int __spu_trap_tag_group(struct spu *spu)
170 pr_debug("%s\n", __FUNCTION__);
171 spu->mfc_callback(spu);
175 static int __spu_trap_spubox(struct spu *spu)
177 if (spu->wbox_callback)
178 spu->wbox_callback(spu);
180 /* atomically disable SPU mailbox interrupts */
181 spin_lock(&spu->register_lock);
182 spu_int_mask_and(spu, 2, ~0x10);
183 spin_unlock(&spu->register_lock);
188 spu_irq_class_0(int irq, void *data, struct pt_regs *regs)
193 spu->class_0_pending = 1;
194 if (spu->stop_callback)
195 spu->stop_callback(spu);
201 spu_irq_class_0_bottom(struct spu *spu)
203 unsigned long stat, mask;
205 spu->class_0_pending = 0;
207 mask = spu_int_mask_get(spu, 0);
208 stat = spu_int_stat_get(spu, 0);
212 if (stat & 1) /* invalid MFC DMA */
213 __spu_trap_invalid_dma(spu);
215 if (stat & 2) /* invalid DMA alignment */
216 __spu_trap_dma_align(spu);
218 if (stat & 4) /* error on SPU */
219 __spu_trap_error(spu);
221 spu_int_stat_clear(spu, 0, stat);
223 return (stat & 0x7) ? -EIO : 0;
225 EXPORT_SYMBOL_GPL(spu_irq_class_0_bottom);
228 spu_irq_class_1(int irq, void *data, struct pt_regs *regs)
231 unsigned long stat, mask, dar, dsisr;
235 /* atomically read & clear class1 status. */
236 spin_lock(&spu->register_lock);
237 mask = spu_int_mask_get(spu, 1);
238 stat = spu_int_stat_get(spu, 1) & mask;
239 dar = spu_mfc_dar_get(spu);
240 dsisr = spu_mfc_dsisr_get(spu);
241 if (stat & 2) /* mapping fault */
242 spu_mfc_dsisr_set(spu, 0ul);
243 spu_int_stat_clear(spu, 1, stat);
244 spin_unlock(&spu->register_lock);
245 pr_debug("%s: %lx %lx %lx %lx\n", __FUNCTION__, mask, stat,
248 if (stat & 1) /* segment fault */
249 __spu_trap_data_seg(spu, dar);
251 if (stat & 2) { /* mapping fault */
252 __spu_trap_data_map(spu, dar, dsisr);
255 if (stat & 4) /* ls compare & suspend on get */
258 if (stat & 8) /* ls compare & suspend on put */
261 return stat ? IRQ_HANDLED : IRQ_NONE;
263 EXPORT_SYMBOL_GPL(spu_irq_class_1_bottom);
266 spu_irq_class_2(int irq, void *data, struct pt_regs *regs)
273 stat = spu_int_stat_get(spu, 2);
274 mask = spu_int_mask_get(spu, 2);
276 pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
280 if (stat & 1) /* PPC core mailbox */
281 __spu_trap_mailbox(spu);
283 if (stat & 2) /* SPU stop-and-signal */
284 __spu_trap_stop(spu);
286 if (stat & 4) /* SPU halted */
287 __spu_trap_halt(spu);
289 if (stat & 8) /* DMA tag group complete */
290 __spu_trap_tag_group(spu);
292 if (stat & 0x10) /* SPU mailbox threshold */
293 __spu_trap_spubox(spu);
295 spu_int_stat_clear(spu, 2, stat);
296 return stat ? IRQ_HANDLED : IRQ_NONE;
300 spu_request_irqs(struct spu *spu)
305 irq_base = IIC_NODE_STRIDE * spu->node + IIC_SPE_OFFSET;
307 snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0", spu->number);
308 ret = request_irq(irq_base + spu->isrc,
309 spu_irq_class_0, 0, spu->irq_c0, spu);
313 snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1", spu->number);
314 ret = request_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc,
315 spu_irq_class_1, 0, spu->irq_c1, spu);
319 snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2", spu->number);
320 ret = request_irq(irq_base + 2*IIC_CLASS_STRIDE + spu->isrc,
321 spu_irq_class_2, 0, spu->irq_c2, spu);
327 free_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc, spu);
329 free_irq(irq_base + spu->isrc, spu);
335 spu_free_irqs(struct spu *spu)
339 irq_base = IIC_NODE_STRIDE * spu->node + IIC_SPE_OFFSET;
341 free_irq(irq_base + spu->isrc, spu);
342 free_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc, spu);
343 free_irq(irq_base + 2*IIC_CLASS_STRIDE + spu->isrc, spu);
346 static LIST_HEAD(spu_list);
347 static DEFINE_MUTEX(spu_mutex);
349 static void spu_init_channels(struct spu *spu)
351 static const struct {
355 { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
356 { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
358 { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
359 { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
360 { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
362 struct spu_priv2 __iomem *priv2;
367 /* initialize all channel data to zero */
368 for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
371 out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
372 for (count = 0; count < zero_list[i].count; count++)
373 out_be64(&priv2->spu_chnldata_RW, 0);
376 /* initialize channel counts to meaningful values */
377 for (i = 0; i < ARRAY_SIZE(count_list); i++) {
378 out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
379 out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
383 struct spu *spu_alloc(void)
387 mutex_lock(&spu_mutex);
388 if (!list_empty(&spu_list)) {
389 spu = list_entry(spu_list.next, struct spu, list);
390 list_del_init(&spu->list);
391 pr_debug("Got SPU %x %d\n", spu->isrc, spu->number);
393 pr_debug("No SPU left\n");
396 mutex_unlock(&spu_mutex);
399 spu_init_channels(spu);
403 EXPORT_SYMBOL_GPL(spu_alloc);
405 void spu_free(struct spu *spu)
407 mutex_lock(&spu_mutex);
408 list_add_tail(&spu->list, &spu_list);
409 mutex_unlock(&spu_mutex);
411 EXPORT_SYMBOL_GPL(spu_free);
413 static int spu_handle_mm_fault(struct spu *spu)
415 struct mm_struct *mm = spu->mm;
416 struct vm_area_struct *vma;
417 u64 ea, dsisr, is_write;
423 if (!IS_VALID_EA(ea)) {
430 if (mm->pgd == NULL) {
434 down_read(&mm->mmap_sem);
435 vma = find_vma(mm, ea);
438 if (vma->vm_start <= ea)
440 if (!(vma->vm_flags & VM_GROWSDOWN))
443 if (expand_stack(vma, ea))
447 is_write = dsisr & MFC_DSISR_ACCESS_PUT;
449 if (!(vma->vm_flags & VM_WRITE))
452 if (dsisr & MFC_DSISR_ACCESS_DENIED)
454 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
458 switch (handle_mm_fault(mm, vma, ea, is_write)) {
465 case VM_FAULT_SIGBUS:
474 up_read(&mm->mmap_sem);
478 up_read(&mm->mmap_sem);
482 int spu_irq_class_1_bottom(struct spu *spu)
484 u64 ea, dsisr, access, error = 0UL;
489 if (dsisr & (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED)) {
490 access = (_PAGE_PRESENT | _PAGE_USER);
491 access |= (dsisr & MFC_DSISR_ACCESS_PUT) ? _PAGE_RW : 0UL;
492 if (hash_page(ea, access, 0x300) != 0)
493 error |= CLASS1_ENABLE_STORAGE_FAULT_INTR;
495 if (error & CLASS1_ENABLE_STORAGE_FAULT_INTR) {
496 if ((ret = spu_handle_mm_fault(spu)) != 0)
497 error |= CLASS1_ENABLE_STORAGE_FAULT_INTR;
499 error &= ~CLASS1_ENABLE_STORAGE_FAULT_INTR;
504 spu_restart_dma(spu);
506 __spu_trap_invalid_dma(spu);
511 void spu_irq_setaffinity(struct spu *spu, int cpu)
513 u64 target = iic_get_target_id(cpu);
514 u64 route = target << 48 | target << 32 | target << 16;
515 spu_int_route_set(spu, route);
517 EXPORT_SYMBOL_GPL(spu_irq_setaffinity);
519 static void __iomem * __init map_spe_prop(struct device_node *n,
522 struct address_prop {
523 unsigned long address;
525 } __attribute__((packed)) *prop;
530 p = get_property(n, name, &proplen);
531 if (proplen != sizeof (struct address_prop))
536 return ioremap(prop->address, prop->len);
539 static void spu_unmap(struct spu *spu)
543 iounmap(spu->problem);
544 iounmap((u8 __iomem *)spu->local_store);
547 static int __init spu_map_device(struct spu *spu, struct device_node *spe)
553 prop = get_property(spe, "isrc", NULL);
556 spu->isrc = *(unsigned int *)prop;
558 spu->name = get_property(spe, "name", NULL);
562 prop = get_property(spe, "local-store", NULL);
565 spu->local_store_phys = *(unsigned long *)prop;
567 /* we use local store as ram, not io memory */
568 spu->local_store = (void __force *)map_spe_prop(spe, "local-store");
569 if (!spu->local_store)
572 prop = get_property(spe, "problem", NULL);
575 spu->problem_phys = *(unsigned long *)prop;
577 spu->problem= map_spe_prop(spe, "problem");
581 spu->priv1= map_spe_prop(spe, "priv1");
582 /* priv1 is not available on a hypervisor */
584 spu->priv2= map_spe_prop(spe, "priv2");
596 static int __init find_spu_node_id(struct device_node *spe)
599 struct device_node *cpu;
601 cpu = spe->parent->parent;
602 id = (unsigned int *)get_property(cpu, "node-id", NULL);
607 static int __init create_spu(struct device_node *spe)
614 spu = kmalloc(sizeof (*spu), GFP_KERNEL);
618 ret = spu_map_device(spu, spe);
622 spu->node = find_spu_node_id(spe);
624 spu->slb_replace = 0;
629 spu->class_0_pending = 0;
633 spin_lock_init(&spu->register_lock);
635 spu_mfc_sdr_set(spu, mfspr(SPRN_SDR1));
636 spu_mfc_sr1_set(spu, 0x33);
638 spu->ibox_callback = NULL;
639 spu->wbox_callback = NULL;
640 spu->stop_callback = NULL;
641 spu->mfc_callback = NULL;
643 mutex_lock(&spu_mutex);
644 spu->number = number++;
645 ret = spu_request_irqs(spu);
649 list_add(&spu->list, &spu_list);
650 mutex_unlock(&spu_mutex);
652 pr_debug(KERN_DEBUG "Using SPE %s %02x %p %p %p %p %d\n",
653 spu->name, spu->isrc, spu->local_store,
654 spu->problem, spu->priv1, spu->priv2, spu->number);
658 mutex_unlock(&spu_mutex);
666 static void destroy_spu(struct spu *spu)
668 list_del_init(&spu->list);
675 static void cleanup_spu_base(void)
677 struct spu *spu, *tmp;
678 mutex_lock(&spu_mutex);
679 list_for_each_entry_safe(spu, tmp, &spu_list, list)
681 mutex_unlock(&spu_mutex);
683 module_exit(cleanup_spu_base);
685 static int __init init_spu_base(void)
687 struct device_node *node;
691 for (node = of_find_node_by_type(NULL, "spe");
692 node; node = of_find_node_by_type(node, "spe")) {
693 ret = create_spu(node);
695 printk(KERN_WARNING "%s: Error initializing %s\n",
696 __FUNCTION__, node->name);
701 /* in some old firmware versions, the spe is called 'spc', so we
702 look for that as well */
703 for (node = of_find_node_by_type(NULL, "spc");
704 node; node = of_find_node_by_type(node, "spc")) {
705 ret = create_spu(node);
707 printk(KERN_WARNING "%s: Error initializing %s\n",
708 __FUNCTION__, node->name);
715 module_init(init_spu_base);
717 MODULE_LICENSE("GPL");
718 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");