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;
85 if (spu_mfc_sr1_get(spu) & MFC_STATE1_RELOCATE_MASK)
86 out_be64(&priv2->slb_invalidate_all_W, 0UL);
88 EXPORT_SYMBOL_GPL(spu_invalidate_slbs);
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
93 void spu_flush_all_slbs(struct mm_struct *mm)
98 spin_lock_irqsave(&spu_full_list_lock, flags);
99 list_for_each_entry(spu, &spu_full_list, full_list) {
101 spu_invalidate_slbs(spu);
103 spin_unlock_irqrestore(&spu_full_list_lock, flags);
106 /* The hack below stinks... try to do something better one of
107 * these days... Does it even work properly with NR_CPUS == 1 ?
109 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
111 int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
113 /* Global TLBIE broadcast required with SPEs. */
114 __cpus_setall(&mm->cpu_vm_mask, nr);
117 void spu_associate_mm(struct spu *spu, struct mm_struct *mm)
121 spin_lock_irqsave(&spu_full_list_lock, flags);
123 spin_unlock_irqrestore(&spu_full_list_lock, flags);
125 mm_needs_global_tlbie(mm);
127 EXPORT_SYMBOL_GPL(spu_associate_mm);
129 int spu_64k_pages_available(void)
131 return mmu_psize_defs[MMU_PAGE_64K].shift != 0;
133 EXPORT_SYMBOL_GPL(spu_64k_pages_available);
135 static void spu_restart_dma(struct spu *spu)
137 struct spu_priv2 __iomem *priv2 = spu->priv2;
139 if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
140 out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
143 static inline void spu_load_slb(struct spu *spu, int slbe, struct spu_slb *slb)
145 struct spu_priv2 __iomem *priv2 = spu->priv2;
147 pr_debug("%s: adding SLB[%d] 0x%016lx 0x%016lx\n",
148 __func__, slbe, slb->vsid, slb->esid);
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);
155 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
157 struct mm_struct *mm = spu->mm;
161 pr_debug("%s\n", __FUNCTION__);
163 slb.esid = (ea & ESID_MASK) | SLB_ESID_V;
165 switch(REGION_ID(ea)) {
167 #ifdef CONFIG_PPC_MM_SLICES
168 psize = get_slice_psize(mm, ea);
170 psize = mm->context.user_psize;
172 slb.vsid = (get_vsid(mm->context.id, ea, MMU_SEGSIZE_256M)
173 << SLB_VSID_SHIFT) | SLB_VSID_USER;
175 case VMALLOC_REGION_ID:
176 if (ea < VMALLOC_END)
177 psize = mmu_vmalloc_psize;
179 psize = mmu_io_psize;
180 slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
181 << SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
183 case KERNEL_REGION_ID:
184 psize = mmu_linear_psize;
185 slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
186 << SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
189 /* Future: support kernel segments so that drivers
192 pr_debug("invalid region access at %016lx\n", ea);
195 slb.vsid |= mmu_psize_defs[psize].sllp;
197 spu_load_slb(spu, spu->slb_replace, &slb);
200 if (spu->slb_replace >= 8)
201 spu->slb_replace = 0;
203 spu_restart_dma(spu);
204 spu->stats.slb_flt++;
208 extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX
209 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
211 pr_debug("%s, %lx, %lx\n", __FUNCTION__, dsisr, ea);
213 /* Handle kernel space hash faults immediately.
214 User hash faults need to be deferred to process context. */
215 if ((dsisr & MFC_DSISR_PTE_NOT_FOUND)
216 && REGION_ID(ea) != USER_REGION_ID
217 && hash_page(ea, _PAGE_PRESENT, 0x300) == 0) {
218 spu_restart_dma(spu);
222 spu->class_0_pending = 0;
226 spu->stop_callback(spu);
231 static void __spu_kernel_slb(void *addr, struct spu_slb *slb)
233 unsigned long ea = (unsigned long)addr;
236 if (REGION_ID(ea) == KERNEL_REGION_ID)
237 llp = mmu_psize_defs[mmu_linear_psize].sllp;
239 llp = mmu_psize_defs[mmu_virtual_psize].sllp;
241 slb->vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M) << SLB_VSID_SHIFT) |
242 SLB_VSID_KERNEL | llp;
243 slb->esid = (ea & ESID_MASK) | SLB_ESID_V;
247 * Given an array of @nr_slbs SLB entries, @slbs, return non-zero if the
248 * address @new_addr is present.
250 static inline int __slb_present(struct spu_slb *slbs, int nr_slbs,
253 unsigned long ea = (unsigned long)new_addr;
256 for (i = 0; i < nr_slbs; i++)
257 if (!((slbs[i].esid ^ ea) & ESID_MASK))
264 * Setup the SPU kernel SLBs, in preparation for a context save/restore. We
265 * need to map both the context save area, and the save/restore code.
267 * Because the lscsa and code may cross segment boundaires, we check to see
268 * if mappings are required for the start and end of each range. We currently
269 * assume that the mappings are smaller that one segment - if not, something
270 * is seriously wrong.
272 void spu_setup_kernel_slbs(struct spu *spu, struct spu_lscsa *lscsa,
273 void *code, int code_size)
275 struct spu_slb slbs[4];
277 /* start and end addresses of both mappings */
279 lscsa, (void *)lscsa + sizeof(*lscsa) - 1,
280 code, code + code_size - 1
283 /* check the set of addresses, and create a new entry in the slbs array
284 * if there isn't already a SLB for that address */
285 for (i = 0; i < ARRAY_SIZE(addrs); i++) {
286 if (__slb_present(slbs, nr_slbs, addrs[i]))
289 __spu_kernel_slb(addrs[i], &slbs[nr_slbs]);
293 /* Add the set of SLBs */
294 for (i = 0; i < nr_slbs; i++)
295 spu_load_slb(spu, i, &slbs[i]);
297 EXPORT_SYMBOL_GPL(spu_setup_kernel_slbs);
300 spu_irq_class_0(int irq, void *data)
303 unsigned long stat, mask;
307 spin_lock(&spu->register_lock);
308 mask = spu_int_mask_get(spu, 0);
309 stat = spu_int_stat_get(spu, 0) & mask;
311 spu->class_0_pending |= stat;
312 spu->dsisr = spu_mfc_dsisr_get(spu);
313 spu->dar = spu_mfc_dar_get(spu);
314 spin_unlock(&spu->register_lock);
316 spu->stop_callback(spu);
318 spu_int_stat_clear(spu, 0, stat);
324 spu_irq_class_1(int irq, void *data)
327 unsigned long stat, mask, dar, dsisr;
331 /* atomically read & clear class1 status. */
332 spin_lock(&spu->register_lock);
333 mask = spu_int_mask_get(spu, 1);
334 stat = spu_int_stat_get(spu, 1) & mask;
335 dar = spu_mfc_dar_get(spu);
336 dsisr = spu_mfc_dsisr_get(spu);
337 if (stat & CLASS1_STORAGE_FAULT_INTR)
338 spu_mfc_dsisr_set(spu, 0ul);
339 spu_int_stat_clear(spu, 1, stat);
340 spin_unlock(&spu->register_lock);
341 pr_debug("%s: %lx %lx %lx %lx\n", __FUNCTION__, mask, stat,
344 if (stat & CLASS1_SEGMENT_FAULT_INTR)
345 __spu_trap_data_seg(spu, dar);
347 if (stat & CLASS1_STORAGE_FAULT_INTR)
348 __spu_trap_data_map(spu, dar, dsisr);
350 if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_GET_INTR)
353 if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_PUT_INTR)
356 return stat ? IRQ_HANDLED : IRQ_NONE;
360 spu_irq_class_2(int irq, void *data)
365 const int mailbox_intrs =
366 CLASS2_MAILBOX_THRESHOLD_INTR | CLASS2_MAILBOX_INTR;
369 spin_lock(&spu->register_lock);
370 stat = spu_int_stat_get(spu, 2);
371 mask = spu_int_mask_get(spu, 2);
372 /* ignore interrupts we're not waiting for */
375 /* mailbox interrupts are level triggered. mask them now before
377 if (stat & mailbox_intrs)
378 spu_int_mask_and(spu, 2, ~(stat & mailbox_intrs));
379 /* acknowledge all interrupts before the callbacks */
380 spu_int_stat_clear(spu, 2, stat);
381 spin_unlock(&spu->register_lock);
383 pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
385 if (stat & CLASS2_MAILBOX_INTR)
386 spu->ibox_callback(spu);
388 if (stat & CLASS2_SPU_STOP_INTR)
389 spu->stop_callback(spu);
391 if (stat & CLASS2_SPU_HALT_INTR)
392 spu->stop_callback(spu);
394 if (stat & CLASS2_SPU_DMA_TAG_GROUP_COMPLETE_INTR)
395 spu->mfc_callback(spu);
397 if (stat & CLASS2_MAILBOX_THRESHOLD_INTR)
398 spu->wbox_callback(spu);
400 spu->stats.class2_intr++;
401 return stat ? IRQ_HANDLED : IRQ_NONE;
404 static int spu_request_irqs(struct spu *spu)
408 if (spu->irqs[0] != NO_IRQ) {
409 snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0",
411 ret = request_irq(spu->irqs[0], spu_irq_class_0,
417 if (spu->irqs[1] != NO_IRQ) {
418 snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1",
420 ret = request_irq(spu->irqs[1], spu_irq_class_1,
426 if (spu->irqs[2] != NO_IRQ) {
427 snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2",
429 ret = request_irq(spu->irqs[2], spu_irq_class_2,
438 if (spu->irqs[1] != NO_IRQ)
439 free_irq(spu->irqs[1], spu);
441 if (spu->irqs[0] != NO_IRQ)
442 free_irq(spu->irqs[0], spu);
447 static void spu_free_irqs(struct spu *spu)
449 if (spu->irqs[0] != NO_IRQ)
450 free_irq(spu->irqs[0], spu);
451 if (spu->irqs[1] != NO_IRQ)
452 free_irq(spu->irqs[1], spu);
453 if (spu->irqs[2] != NO_IRQ)
454 free_irq(spu->irqs[2], spu);
457 void spu_init_channels(struct spu *spu)
459 static const struct {
463 { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
464 { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
466 { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
467 { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
468 { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
470 struct spu_priv2 __iomem *priv2;
475 /* initialize all channel data to zero */
476 for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
479 out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
480 for (count = 0; count < zero_list[i].count; count++)
481 out_be64(&priv2->spu_chnldata_RW, 0);
484 /* initialize channel counts to meaningful values */
485 for (i = 0; i < ARRAY_SIZE(count_list); i++) {
486 out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
487 out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
490 EXPORT_SYMBOL_GPL(spu_init_channels);
492 static int spu_shutdown(struct sys_device *sysdev)
494 struct spu *spu = container_of(sysdev, struct spu, sysdev);
497 spu_destroy_spu(spu);
501 static struct sysdev_class spu_sysdev_class = {
503 .shutdown = spu_shutdown,
506 int spu_add_sysdev_attr(struct sysdev_attribute *attr)
510 mutex_lock(&spu_full_list_mutex);
511 list_for_each_entry(spu, &spu_full_list, full_list)
512 sysdev_create_file(&spu->sysdev, attr);
513 mutex_unlock(&spu_full_list_mutex);
517 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr);
519 int spu_add_sysdev_attr_group(struct attribute_group *attrs)
524 mutex_lock(&spu_full_list_mutex);
525 list_for_each_entry(spu, &spu_full_list, full_list) {
526 rc = sysfs_create_group(&spu->sysdev.kobj, attrs);
528 /* we're in trouble here, but try unwinding anyway */
530 printk(KERN_ERR "%s: can't create sysfs group '%s'\n",
531 __func__, attrs->name);
533 list_for_each_entry_continue_reverse(spu,
534 &spu_full_list, full_list)
535 sysfs_remove_group(&spu->sysdev.kobj, attrs);
540 mutex_unlock(&spu_full_list_mutex);
544 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr_group);
547 void spu_remove_sysdev_attr(struct sysdev_attribute *attr)
551 mutex_lock(&spu_full_list_mutex);
552 list_for_each_entry(spu, &spu_full_list, full_list)
553 sysdev_remove_file(&spu->sysdev, attr);
554 mutex_unlock(&spu_full_list_mutex);
556 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr);
558 void spu_remove_sysdev_attr_group(struct attribute_group *attrs)
562 mutex_lock(&spu_full_list_mutex);
563 list_for_each_entry(spu, &spu_full_list, full_list)
564 sysfs_remove_group(&spu->sysdev.kobj, attrs);
565 mutex_unlock(&spu_full_list_mutex);
567 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr_group);
569 static int spu_create_sysdev(struct spu *spu)
573 spu->sysdev.id = spu->number;
574 spu->sysdev.cls = &spu_sysdev_class;
575 ret = sysdev_register(&spu->sysdev);
577 printk(KERN_ERR "Can't register SPU %d with sysfs\n",
582 sysfs_add_device_to_node(&spu->sysdev, spu->node);
587 static int __init create_spu(void *data)
596 spu = kzalloc(sizeof (*spu), GFP_KERNEL);
600 spu->alloc_state = SPU_FREE;
602 spin_lock_init(&spu->register_lock);
603 spin_lock(&spu_lock);
604 spu->number = number++;
605 spin_unlock(&spu_lock);
607 ret = spu_create_spu(spu, data);
612 spu_mfc_sdr_setup(spu);
613 spu_mfc_sr1_set(spu, 0x33);
614 ret = spu_request_irqs(spu);
618 ret = spu_create_sysdev(spu);
622 mutex_lock(&cbe_spu_info[spu->node].list_mutex);
623 list_add(&spu->cbe_list, &cbe_spu_info[spu->node].spus);
624 cbe_spu_info[spu->node].n_spus++;
625 mutex_unlock(&cbe_spu_info[spu->node].list_mutex);
627 mutex_lock(&spu_full_list_mutex);
628 spin_lock_irqsave(&spu_full_list_lock, flags);
629 list_add(&spu->full_list, &spu_full_list);
630 spin_unlock_irqrestore(&spu_full_list_lock, flags);
631 mutex_unlock(&spu_full_list_mutex);
633 spu->stats.util_state = SPU_UTIL_IDLE_LOADED;
635 spu->stats.tstamp = timespec_to_ns(&ts);
637 INIT_LIST_HEAD(&spu->aff_list);
644 spu_destroy_spu(spu);
651 static const char *spu_state_names[] = {
652 "user", "system", "iowait", "idle"
655 static unsigned long long spu_acct_time(struct spu *spu,
656 enum spu_utilization_state state)
659 unsigned long long time = spu->stats.times[state];
662 * If the spu is idle or the context is stopped, utilization
663 * statistics are not updated. Apply the time delta from the
664 * last recorded state of the spu.
666 if (spu->stats.util_state == state) {
668 time += timespec_to_ns(&ts) - spu->stats.tstamp;
671 return time / NSEC_PER_MSEC;
675 static ssize_t spu_stat_show(struct sys_device *sysdev, char *buf)
677 struct spu *spu = container_of(sysdev, struct spu, sysdev);
679 return sprintf(buf, "%s %llu %llu %llu %llu "
680 "%llu %llu %llu %llu %llu %llu %llu %llu\n",
681 spu_state_names[spu->stats.util_state],
682 spu_acct_time(spu, SPU_UTIL_USER),
683 spu_acct_time(spu, SPU_UTIL_SYSTEM),
684 spu_acct_time(spu, SPU_UTIL_IOWAIT),
685 spu_acct_time(spu, SPU_UTIL_IDLE_LOADED),
686 spu->stats.vol_ctx_switch,
687 spu->stats.invol_ctx_switch,
692 spu->stats.class2_intr,
693 spu->stats.libassist);
696 static SYSDEV_ATTR(stat, 0644, spu_stat_show, NULL);
698 static int __init init_spu_base(void)
702 for (i = 0; i < MAX_NUMNODES; i++) {
703 mutex_init(&cbe_spu_info[i].list_mutex);
704 INIT_LIST_HEAD(&cbe_spu_info[i].spus);
707 if (!spu_management_ops)
710 /* create sysdev class for spus */
711 ret = sysdev_class_register(&spu_sysdev_class);
715 ret = spu_enumerate_spus(create_spu);
718 printk(KERN_WARNING "%s: Error initializing spus\n",
720 goto out_unregister_sysdev_class;
725 * We cannot put the forward declaration in
726 * <linux/linux_logo.h> because of conflicting session type
727 * conflicts for const and __initdata with different compiler
730 extern const struct linux_logo logo_spe_clut224;
732 fb_append_extra_logo(&logo_spe_clut224, ret);
735 mutex_lock(&spu_full_list_mutex);
736 xmon_register_spus(&spu_full_list);
737 crash_register_spus(&spu_full_list);
738 mutex_unlock(&spu_full_list_mutex);
739 spu_add_sysdev_attr(&attr_stat);
745 out_unregister_sysdev_class:
746 sysdev_class_unregister(&spu_sysdev_class);
750 module_init(init_spu_base);
752 MODULE_LICENSE("GPL");
753 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");