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", __FUNCTION__
);
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", __FUNCTION__
, 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_0_pending
= 0;
233 spu
->stop_callback(spu
);
238 static void __spu_kernel_slb(void *addr
, struct spu_slb
*slb
)
240 unsigned long ea
= (unsigned long)addr
;
243 if (REGION_ID(ea
) == KERNEL_REGION_ID
)
244 llp
= mmu_psize_defs
[mmu_linear_psize
].sllp
;
246 llp
= mmu_psize_defs
[mmu_virtual_psize
].sllp
;
248 slb
->vsid
= (get_kernel_vsid(ea
, MMU_SEGSIZE_256M
) << SLB_VSID_SHIFT
) |
249 SLB_VSID_KERNEL
| llp
;
250 slb
->esid
= (ea
& ESID_MASK
) | SLB_ESID_V
;
254 * Given an array of @nr_slbs SLB entries, @slbs, return non-zero if the
255 * address @new_addr is present.
257 static inline int __slb_present(struct spu_slb
*slbs
, int nr_slbs
,
260 unsigned long ea
= (unsigned long)new_addr
;
263 for (i
= 0; i
< nr_slbs
; i
++)
264 if (!((slbs
[i
].esid
^ ea
) & ESID_MASK
))
271 * Setup the SPU kernel SLBs, in preparation for a context save/restore. We
272 * need to map both the context save area, and the save/restore code.
274 * Because the lscsa and code may cross segment boundaires, we check to see
275 * if mappings are required for the start and end of each range. We currently
276 * assume that the mappings are smaller that one segment - if not, something
277 * is seriously wrong.
279 void spu_setup_kernel_slbs(struct spu
*spu
, struct spu_lscsa
*lscsa
,
280 void *code
, int code_size
)
282 struct spu_slb slbs
[4];
284 /* start and end addresses of both mappings */
286 lscsa
, (void *)lscsa
+ sizeof(*lscsa
) - 1,
287 code
, code
+ code_size
- 1
290 /* check the set of addresses, and create a new entry in the slbs array
291 * if there isn't already a SLB for that address */
292 for (i
= 0; i
< ARRAY_SIZE(addrs
); i
++) {
293 if (__slb_present(slbs
, nr_slbs
, addrs
[i
]))
296 __spu_kernel_slb(addrs
[i
], &slbs
[nr_slbs
]);
300 spin_lock_irq(&spu
->register_lock
);
301 /* Add the set of SLBs */
302 for (i
= 0; i
< nr_slbs
; i
++)
303 spu_load_slb(spu
, i
, &slbs
[i
]);
304 spin_unlock_irq(&spu
->register_lock
);
306 EXPORT_SYMBOL_GPL(spu_setup_kernel_slbs
);
309 spu_irq_class_0(int irq
, void *data
)
312 unsigned long stat
, mask
;
316 spin_lock(&spu
->register_lock
);
317 mask
= spu_int_mask_get(spu
, 0);
318 stat
= spu_int_stat_get(spu
, 0) & mask
;
320 spu
->class_0_pending
|= stat
;
321 spu
->dsisr
= spu_mfc_dsisr_get(spu
);
322 spu
->dar
= spu_mfc_dar_get(spu
);
323 spin_unlock(&spu
->register_lock
);
325 spu
->stop_callback(spu
);
327 spu_int_stat_clear(spu
, 0, stat
);
333 spu_irq_class_1(int irq
, void *data
)
336 unsigned long stat
, mask
, dar
, dsisr
;
340 /* atomically read & clear class1 status. */
341 spin_lock(&spu
->register_lock
);
342 mask
= spu_int_mask_get(spu
, 1);
343 stat
= spu_int_stat_get(spu
, 1) & mask
;
344 dar
= spu_mfc_dar_get(spu
);
345 dsisr
= spu_mfc_dsisr_get(spu
);
346 if (stat
& CLASS1_STORAGE_FAULT_INTR
)
347 spu_mfc_dsisr_set(spu
, 0ul);
348 spu_int_stat_clear(spu
, 1, stat
);
350 if (stat
& CLASS1_SEGMENT_FAULT_INTR
)
351 __spu_trap_data_seg(spu
, dar
);
353 spin_unlock(&spu
->register_lock
);
354 pr_debug("%s: %lx %lx %lx %lx\n", __FUNCTION__
, mask
, stat
,
357 if (stat
& CLASS1_STORAGE_FAULT_INTR
)
358 __spu_trap_data_map(spu
, dar
, dsisr
);
360 if (stat
& CLASS1_LS_COMPARE_SUSPEND_ON_GET_INTR
)
363 if (stat
& CLASS1_LS_COMPARE_SUSPEND_ON_PUT_INTR
)
366 return stat
? IRQ_HANDLED
: IRQ_NONE
;
370 spu_irq_class_2(int irq
, void *data
)
375 const int mailbox_intrs
=
376 CLASS2_MAILBOX_THRESHOLD_INTR
| CLASS2_MAILBOX_INTR
;
379 spin_lock(&spu
->register_lock
);
380 stat
= spu_int_stat_get(spu
, 2);
381 mask
= spu_int_mask_get(spu
, 2);
382 /* ignore interrupts we're not waiting for */
385 /* mailbox interrupts are level triggered. mask them now before
387 if (stat
& mailbox_intrs
)
388 spu_int_mask_and(spu
, 2, ~(stat
& mailbox_intrs
));
389 /* acknowledge all interrupts before the callbacks */
390 spu_int_stat_clear(spu
, 2, stat
);
391 spin_unlock(&spu
->register_lock
);
393 pr_debug("class 2 interrupt %d, %lx, %lx\n", irq
, stat
, mask
);
395 if (stat
& CLASS2_MAILBOX_INTR
)
396 spu
->ibox_callback(spu
);
398 if (stat
& CLASS2_SPU_STOP_INTR
)
399 spu
->stop_callback(spu
);
401 if (stat
& CLASS2_SPU_HALT_INTR
)
402 spu
->stop_callback(spu
);
404 if (stat
& CLASS2_SPU_DMA_TAG_GROUP_COMPLETE_INTR
)
405 spu
->mfc_callback(spu
);
407 if (stat
& CLASS2_MAILBOX_THRESHOLD_INTR
)
408 spu
->wbox_callback(spu
);
410 spu
->stats
.class2_intr
++;
411 return stat
? IRQ_HANDLED
: IRQ_NONE
;
414 static int spu_request_irqs(struct spu
*spu
)
418 if (spu
->irqs
[0] != NO_IRQ
) {
419 snprintf(spu
->irq_c0
, sizeof (spu
->irq_c0
), "spe%02d.0",
421 ret
= request_irq(spu
->irqs
[0], spu_irq_class_0
,
427 if (spu
->irqs
[1] != NO_IRQ
) {
428 snprintf(spu
->irq_c1
, sizeof (spu
->irq_c1
), "spe%02d.1",
430 ret
= request_irq(spu
->irqs
[1], spu_irq_class_1
,
436 if (spu
->irqs
[2] != NO_IRQ
) {
437 snprintf(spu
->irq_c2
, sizeof (spu
->irq_c2
), "spe%02d.2",
439 ret
= request_irq(spu
->irqs
[2], spu_irq_class_2
,
448 if (spu
->irqs
[1] != NO_IRQ
)
449 free_irq(spu
->irqs
[1], spu
);
451 if (spu
->irqs
[0] != NO_IRQ
)
452 free_irq(spu
->irqs
[0], spu
);
457 static void spu_free_irqs(struct spu
*spu
)
459 if (spu
->irqs
[0] != NO_IRQ
)
460 free_irq(spu
->irqs
[0], spu
);
461 if (spu
->irqs
[1] != NO_IRQ
)
462 free_irq(spu
->irqs
[1], spu
);
463 if (spu
->irqs
[2] != NO_IRQ
)
464 free_irq(spu
->irqs
[2], spu
);
467 void spu_init_channels(struct spu
*spu
)
469 static const struct {
473 { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
474 { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
476 { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
477 { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
478 { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
480 struct spu_priv2 __iomem
*priv2
;
485 /* initialize all channel data to zero */
486 for (i
= 0; i
< ARRAY_SIZE(zero_list
); i
++) {
489 out_be64(&priv2
->spu_chnlcntptr_RW
, zero_list
[i
].channel
);
490 for (count
= 0; count
< zero_list
[i
].count
; count
++)
491 out_be64(&priv2
->spu_chnldata_RW
, 0);
494 /* initialize channel counts to meaningful values */
495 for (i
= 0; i
< ARRAY_SIZE(count_list
); i
++) {
496 out_be64(&priv2
->spu_chnlcntptr_RW
, count_list
[i
].channel
);
497 out_be64(&priv2
->spu_chnlcnt_RW
, count_list
[i
].count
);
500 EXPORT_SYMBOL_GPL(spu_init_channels
);
502 static int spu_shutdown(struct sys_device
*sysdev
)
504 struct spu
*spu
= container_of(sysdev
, struct spu
, sysdev
);
507 spu_destroy_spu(spu
);
511 static struct sysdev_class spu_sysdev_class
= {
513 .shutdown
= spu_shutdown
,
516 int spu_add_sysdev_attr(struct sysdev_attribute
*attr
)
520 mutex_lock(&spu_full_list_mutex
);
521 list_for_each_entry(spu
, &spu_full_list
, full_list
)
522 sysdev_create_file(&spu
->sysdev
, attr
);
523 mutex_unlock(&spu_full_list_mutex
);
527 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr
);
529 int spu_add_sysdev_attr_group(struct attribute_group
*attrs
)
534 mutex_lock(&spu_full_list_mutex
);
535 list_for_each_entry(spu
, &spu_full_list
, full_list
) {
536 rc
= sysfs_create_group(&spu
->sysdev
.kobj
, attrs
);
538 /* we're in trouble here, but try unwinding anyway */
540 printk(KERN_ERR
"%s: can't create sysfs group '%s'\n",
541 __func__
, attrs
->name
);
543 list_for_each_entry_continue_reverse(spu
,
544 &spu_full_list
, full_list
)
545 sysfs_remove_group(&spu
->sysdev
.kobj
, attrs
);
550 mutex_unlock(&spu_full_list_mutex
);
554 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr_group
);
557 void spu_remove_sysdev_attr(struct sysdev_attribute
*attr
)
561 mutex_lock(&spu_full_list_mutex
);
562 list_for_each_entry(spu
, &spu_full_list
, full_list
)
563 sysdev_remove_file(&spu
->sysdev
, attr
);
564 mutex_unlock(&spu_full_list_mutex
);
566 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr
);
568 void spu_remove_sysdev_attr_group(struct attribute_group
*attrs
)
572 mutex_lock(&spu_full_list_mutex
);
573 list_for_each_entry(spu
, &spu_full_list
, full_list
)
574 sysfs_remove_group(&spu
->sysdev
.kobj
, attrs
);
575 mutex_unlock(&spu_full_list_mutex
);
577 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr_group
);
579 static int spu_create_sysdev(struct spu
*spu
)
583 spu
->sysdev
.id
= spu
->number
;
584 spu
->sysdev
.cls
= &spu_sysdev_class
;
585 ret
= sysdev_register(&spu
->sysdev
);
587 printk(KERN_ERR
"Can't register SPU %d with sysfs\n",
592 sysfs_add_device_to_node(&spu
->sysdev
, spu
->node
);
597 static int __init
create_spu(void *data
)
606 spu
= kzalloc(sizeof (*spu
), GFP_KERNEL
);
610 spu
->alloc_state
= SPU_FREE
;
612 spin_lock_init(&spu
->register_lock
);
613 spin_lock(&spu_lock
);
614 spu
->number
= number
++;
615 spin_unlock(&spu_lock
);
617 ret
= spu_create_spu(spu
, data
);
622 spu_mfc_sdr_setup(spu
);
623 spu_mfc_sr1_set(spu
, 0x33);
624 ret
= spu_request_irqs(spu
);
628 ret
= spu_create_sysdev(spu
);
632 mutex_lock(&cbe_spu_info
[spu
->node
].list_mutex
);
633 list_add(&spu
->cbe_list
, &cbe_spu_info
[spu
->node
].spus
);
634 cbe_spu_info
[spu
->node
].n_spus
++;
635 mutex_unlock(&cbe_spu_info
[spu
->node
].list_mutex
);
637 mutex_lock(&spu_full_list_mutex
);
638 spin_lock_irqsave(&spu_full_list_lock
, flags
);
639 list_add(&spu
->full_list
, &spu_full_list
);
640 spin_unlock_irqrestore(&spu_full_list_lock
, flags
);
641 mutex_unlock(&spu_full_list_mutex
);
643 spu
->stats
.util_state
= SPU_UTIL_IDLE_LOADED
;
645 spu
->stats
.tstamp
= timespec_to_ns(&ts
);
647 INIT_LIST_HEAD(&spu
->aff_list
);
654 spu_destroy_spu(spu
);
661 static const char *spu_state_names
[] = {
662 "user", "system", "iowait", "idle"
665 static unsigned long long spu_acct_time(struct spu
*spu
,
666 enum spu_utilization_state state
)
669 unsigned long long time
= spu
->stats
.times
[state
];
672 * If the spu is idle or the context is stopped, utilization
673 * statistics are not updated. Apply the time delta from the
674 * last recorded state of the spu.
676 if (spu
->stats
.util_state
== state
) {
678 time
+= timespec_to_ns(&ts
) - spu
->stats
.tstamp
;
681 return time
/ NSEC_PER_MSEC
;
685 static ssize_t
spu_stat_show(struct sys_device
*sysdev
, char *buf
)
687 struct spu
*spu
= container_of(sysdev
, struct spu
, sysdev
);
689 return sprintf(buf
, "%s %llu %llu %llu %llu "
690 "%llu %llu %llu %llu %llu %llu %llu %llu\n",
691 spu_state_names
[spu
->stats
.util_state
],
692 spu_acct_time(spu
, SPU_UTIL_USER
),
693 spu_acct_time(spu
, SPU_UTIL_SYSTEM
),
694 spu_acct_time(spu
, SPU_UTIL_IOWAIT
),
695 spu_acct_time(spu
, SPU_UTIL_IDLE_LOADED
),
696 spu
->stats
.vol_ctx_switch
,
697 spu
->stats
.invol_ctx_switch
,
702 spu
->stats
.class2_intr
,
703 spu
->stats
.libassist
);
706 static SYSDEV_ATTR(stat
, 0644, spu_stat_show
, NULL
);
708 static int __init
init_spu_base(void)
712 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
713 mutex_init(&cbe_spu_info
[i
].list_mutex
);
714 INIT_LIST_HEAD(&cbe_spu_info
[i
].spus
);
717 if (!spu_management_ops
)
720 /* create sysdev class for spus */
721 ret
= sysdev_class_register(&spu_sysdev_class
);
725 ret
= spu_enumerate_spus(create_spu
);
728 printk(KERN_WARNING
"%s: Error initializing spus\n",
730 goto out_unregister_sysdev_class
;
735 * We cannot put the forward declaration in
736 * <linux/linux_logo.h> because of conflicting session type
737 * conflicts for const and __initdata with different compiler
740 extern const struct linux_logo logo_spe_clut224
;
742 fb_append_extra_logo(&logo_spe_clut224
, ret
);
745 mutex_lock(&spu_full_list_mutex
);
746 xmon_register_spus(&spu_full_list
);
747 crash_register_spus(&spu_full_list
);
748 mutex_unlock(&spu_full_list_mutex
);
749 spu_add_sysdev_attr(&attr_stat
);
755 out_unregister_sysdev_class
:
756 sysdev_class_unregister(&spu_sysdev_class
);
760 module_init(init_spu_base
);
762 MODULE_LICENSE("GPL");
763 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");