diff options
author | Arnd Bergmann <arnd@arndb.de> | 2005-12-05 22:52:25 -0500 |
---|---|---|
committer | Paul Mackerras <paulus@samba.org> | 2006-01-08 22:52:55 -0500 |
commit | 5110459f181ef1f11200bb3dec61953f08cc49e7 (patch) | |
tree | 73356ce50b3fb5055b4a6f39f237f046615f797d /arch | |
parent | 3b3d22cb84a0bb12f6bbb2b1158972894bec3f21 (diff) |
[PATCH] spufs: Improved SPU preemptability.
This patch makes it easier to preempt an SPU context by
having the scheduler hold ctx->state_sema for much shorter
periods of time.
As part of this restructuring, the control logic for the "run"
operation is moved from arch/ppc64/kernel/spu_base.c to
fs/spufs/file.c. Of course the base retains "bottom half"
handlers for class{0,1} irqs. The new run loop will re-acquire
an SPU if preempted.
From: Mark Nutter <mnutter@us.ibm.com>
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Diffstat (limited to 'arch')
-rw-r--r-- | arch/powerpc/platforms/cell/spu_base.c | 93 | ||||
-rw-r--r-- | arch/powerpc/platforms/cell/spufs/backing_ops.c | 19 | ||||
-rw-r--r-- | arch/powerpc/platforms/cell/spufs/context.c | 5 | ||||
-rw-r--r-- | arch/powerpc/platforms/cell/spufs/file.c | 117 | ||||
-rw-r--r-- | arch/powerpc/platforms/cell/spufs/hw_ops.c | 17 | ||||
-rw-r--r-- | arch/powerpc/platforms/cell/spufs/sched.c | 67 | ||||
-rw-r--r-- | arch/powerpc/platforms/cell/spufs/spufs.h | 5 |
7 files changed, 202 insertions, 121 deletions
diff --git a/arch/powerpc/platforms/cell/spu_base.c b/arch/powerpc/platforms/cell/spu_base.c index 167580ce8691..8abd4bd19665 100644 --- a/arch/powerpc/platforms/cell/spu_base.c +++ b/arch/powerpc/platforms/cell/spu_base.c | |||
@@ -130,7 +130,8 @@ static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr) | |||
130 | spu->dar = ea; | 130 | spu->dar = ea; |
131 | spu->dsisr = dsisr; | 131 | spu->dsisr = dsisr; |
132 | mb(); | 132 | mb(); |
133 | wake_up(&spu->stop_wq); | 133 | if (spu->stop_callback) |
134 | spu->stop_callback(spu); | ||
134 | return 0; | 135 | return 0; |
135 | } | 136 | } |
136 | 137 | ||
@@ -151,7 +152,8 @@ static int __spu_trap_stop(struct spu *spu) | |||
151 | { | 152 | { |
152 | pr_debug("%s\n", __FUNCTION__); | 153 | pr_debug("%s\n", __FUNCTION__); |
153 | spu->stop_code = in_be32(&spu->problem->spu_status_R); | 154 | spu->stop_code = in_be32(&spu->problem->spu_status_R); |
154 | wake_up(&spu->stop_wq); | 155 | if (spu->stop_callback) |
156 | spu->stop_callback(spu); | ||
155 | return 0; | 157 | return 0; |
156 | } | 158 | } |
157 | 159 | ||
@@ -159,7 +161,8 @@ static int __spu_trap_halt(struct spu *spu) | |||
159 | { | 161 | { |
160 | pr_debug("%s\n", __FUNCTION__); | 162 | pr_debug("%s\n", __FUNCTION__); |
161 | spu->stop_code = in_be32(&spu->problem->spu_status_R); | 163 | spu->stop_code = in_be32(&spu->problem->spu_status_R); |
162 | wake_up(&spu->stop_wq); | 164 | if (spu->stop_callback) |
165 | spu->stop_callback(spu); | ||
163 | return 0; | 166 | return 0; |
164 | } | 167 | } |
165 | 168 | ||
@@ -190,12 +193,13 @@ spu_irq_class_0(int irq, void *data, struct pt_regs *regs) | |||
190 | 193 | ||
191 | spu = data; | 194 | spu = data; |
192 | spu->class_0_pending = 1; | 195 | spu->class_0_pending = 1; |
193 | wake_up(&spu->stop_wq); | 196 | if (spu->stop_callback) |
197 | spu->stop_callback(spu); | ||
194 | 198 | ||
195 | return IRQ_HANDLED; | 199 | return IRQ_HANDLED; |
196 | } | 200 | } |
197 | 201 | ||
198 | static int | 202 | int |
199 | spu_irq_class_0_bottom(struct spu *spu) | 203 | spu_irq_class_0_bottom(struct spu *spu) |
200 | { | 204 | { |
201 | unsigned long stat; | 205 | unsigned long stat; |
@@ -214,8 +218,10 @@ spu_irq_class_0_bottom(struct spu *spu) | |||
214 | __spu_trap_error(spu); | 218 | __spu_trap_error(spu); |
215 | 219 | ||
216 | out_be64(&spu->priv1->int_stat_class0_RW, stat); | 220 | out_be64(&spu->priv1->int_stat_class0_RW, stat); |
217 | return 0; | 221 | |
222 | return (stat & 0x7) ? -EIO : 0; | ||
218 | } | 223 | } |
224 | EXPORT_SYMBOL_GPL(spu_irq_class_0_bottom); | ||
219 | 225 | ||
220 | static irqreturn_t | 226 | static irqreturn_t |
221 | spu_irq_class_1(int irq, void *data, struct pt_regs *regs) | 227 | spu_irq_class_1(int irq, void *data, struct pt_regs *regs) |
@@ -250,6 +256,7 @@ spu_irq_class_1(int irq, void *data, struct pt_regs *regs) | |||
250 | 256 | ||
251 | return stat ? IRQ_HANDLED : IRQ_NONE; | 257 | return stat ? IRQ_HANDLED : IRQ_NONE; |
252 | } | 258 | } |
259 | EXPORT_SYMBOL_GPL(spu_irq_class_1_bottom); | ||
253 | 260 | ||
254 | static irqreturn_t | 261 | static irqreturn_t |
255 | spu_irq_class_2(int irq, void *data, struct pt_regs *regs) | 262 | spu_irq_class_2(int irq, void *data, struct pt_regs *regs) |
@@ -478,7 +485,7 @@ bad_area: | |||
478 | return -EFAULT; | 485 | return -EFAULT; |
479 | } | 486 | } |
480 | 487 | ||
481 | static int spu_handle_pte_fault(struct spu *spu) | 488 | int spu_irq_class_1_bottom(struct spu *spu) |
482 | { | 489 | { |
483 | u64 ea, dsisr, access, error = 0UL; | 490 | u64 ea, dsisr, access, error = 0UL; |
484 | int ret = 0; | 491 | int ret = 0; |
@@ -508,76 +515,6 @@ static int spu_handle_pte_fault(struct spu *spu) | |||
508 | return ret; | 515 | return ret; |
509 | } | 516 | } |
510 | 517 | ||
511 | static inline int spu_pending(struct spu *spu, u32 * stat) | ||
512 | { | ||
513 | struct spu_problem __iomem *prob = spu->problem; | ||
514 | u64 pte_fault; | ||
515 | |||
516 | *stat = in_be32(&prob->spu_status_R); | ||
517 | pte_fault = spu->dsisr & | ||
518 | (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED); | ||
519 | return (!(*stat & 0x1) || pte_fault || spu->class_0_pending) ? 1 : 0; | ||
520 | } | ||
521 | |||
522 | int spu_run(struct spu *spu) | ||
523 | { | ||
524 | struct spu_problem __iomem *prob; | ||
525 | struct spu_priv1 __iomem *priv1; | ||
526 | struct spu_priv2 __iomem *priv2; | ||
527 | u32 status; | ||
528 | int ret; | ||
529 | |||
530 | prob = spu->problem; | ||
531 | priv1 = spu->priv1; | ||
532 | priv2 = spu->priv2; | ||
533 | |||
534 | /* Let SPU run. */ | ||
535 | eieio(); | ||
536 | out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_RUNNABLE); | ||
537 | |||
538 | do { | ||
539 | ret = wait_event_interruptible(spu->stop_wq, | ||
540 | spu_pending(spu, &status)); | ||
541 | |||
542 | if (spu->dsisr & | ||
543 | (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED)) | ||
544 | ret = spu_handle_pte_fault(spu); | ||
545 | |||
546 | if (spu->class_0_pending) | ||
547 | spu_irq_class_0_bottom(spu); | ||
548 | |||
549 | if (!ret && signal_pending(current)) | ||
550 | ret = -ERESTARTSYS; | ||
551 | |||
552 | } while (!ret && !(status & | ||
553 | (SPU_STATUS_STOPPED_BY_STOP | | ||
554 | SPU_STATUS_STOPPED_BY_HALT))); | ||
555 | |||
556 | /* Ensure SPU is stopped. */ | ||
557 | out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_STOP); | ||
558 | eieio(); | ||
559 | while (in_be32(&prob->spu_status_R) & SPU_STATUS_RUNNING) | ||
560 | cpu_relax(); | ||
561 | |||
562 | out_be64(&priv2->slb_invalidate_all_W, 0); | ||
563 | out_be64(&priv1->tlb_invalidate_entry_W, 0UL); | ||
564 | eieio(); | ||
565 | |||
566 | /* Check for SPU breakpoint. */ | ||
567 | if (unlikely(current->ptrace & PT_PTRACED)) { | ||
568 | status = in_be32(&prob->spu_status_R); | ||
569 | |||
570 | if ((status & SPU_STATUS_STOPPED_BY_STOP) | ||
571 | && status >> SPU_STOP_STATUS_SHIFT == 0x3fff) { | ||
572 | force_sig(SIGTRAP, current); | ||
573 | ret = -ERESTARTSYS; | ||
574 | } | ||
575 | } | ||
576 | |||
577 | return ret; | ||
578 | } | ||
579 | EXPORT_SYMBOL_GPL(spu_run); | ||
580 | |||
581 | static void __iomem * __init map_spe_prop(struct device_node *n, | 518 | static void __iomem * __init map_spe_prop(struct device_node *n, |
582 | const char *name) | 519 | const char *name) |
583 | { | 520 | { |
@@ -693,9 +630,9 @@ static int __init create_spu(struct device_node *spe) | |||
693 | out_be64(&spu->priv1->mfc_sdr_RW, mfspr(SPRN_SDR1)); | 630 | out_be64(&spu->priv1->mfc_sdr_RW, mfspr(SPRN_SDR1)); |
694 | out_be64(&spu->priv1->mfc_sr1_RW, 0x33); | 631 | out_be64(&spu->priv1->mfc_sr1_RW, 0x33); |
695 | 632 | ||
696 | init_waitqueue_head(&spu->stop_wq); | ||
697 | spu->ibox_callback = NULL; | 633 | spu->ibox_callback = NULL; |
698 | spu->wbox_callback = NULL; | 634 | spu->wbox_callback = NULL; |
635 | spu->stop_callback = NULL; | ||
699 | 636 | ||
700 | down(&spu_mutex); | 637 | down(&spu_mutex); |
701 | spu->number = number++; | 638 | spu->number = number++; |
diff --git a/arch/powerpc/platforms/cell/spufs/backing_ops.c b/arch/powerpc/platforms/cell/spufs/backing_ops.c index caf0984064e2..66567c109965 100644 --- a/arch/powerpc/platforms/cell/spufs/backing_ops.c +++ b/arch/powerpc/platforms/cell/spufs/backing_ops.c | |||
@@ -232,6 +232,23 @@ static char *spu_backing_get_ls(struct spu_context *ctx) | |||
232 | return ctx->csa.lscsa->ls; | 232 | return ctx->csa.lscsa->ls; |
233 | } | 233 | } |
234 | 234 | ||
235 | static void spu_backing_runcntl_write(struct spu_context *ctx, u32 val) | ||
236 | { | ||
237 | spin_lock(&ctx->csa.register_lock); | ||
238 | ctx->csa.prob.spu_runcntl_RW = val; | ||
239 | if (val & SPU_RUNCNTL_RUNNABLE) { | ||
240 | ctx->csa.prob.spu_status_R |= SPU_STATUS_RUNNING; | ||
241 | } else { | ||
242 | ctx->csa.prob.spu_status_R &= ~SPU_STATUS_RUNNING; | ||
243 | } | ||
244 | spin_unlock(&ctx->csa.register_lock); | ||
245 | } | ||
246 | |||
247 | static void spu_backing_runcntl_stop(struct spu_context *ctx) | ||
248 | { | ||
249 | spu_backing_runcntl_write(ctx, SPU_RUNCNTL_STOP); | ||
250 | } | ||
251 | |||
235 | struct spu_context_ops spu_backing_ops = { | 252 | struct spu_context_ops spu_backing_ops = { |
236 | .mbox_read = spu_backing_mbox_read, | 253 | .mbox_read = spu_backing_mbox_read, |
237 | .mbox_stat_read = spu_backing_mbox_stat_read, | 254 | .mbox_stat_read = spu_backing_mbox_stat_read, |
@@ -249,4 +266,6 @@ struct spu_context_ops spu_backing_ops = { | |||
249 | .npc_write = spu_backing_npc_write, | 266 | .npc_write = spu_backing_npc_write, |
250 | .status_read = spu_backing_status_read, | 267 | .status_read = spu_backing_status_read, |
251 | .get_ls = spu_backing_get_ls, | 268 | .get_ls = spu_backing_get_ls, |
269 | .runcntl_write = spu_backing_runcntl_write, | ||
270 | .runcntl_stop = spu_backing_runcntl_stop, | ||
252 | }; | 271 | }; |
diff --git a/arch/powerpc/platforms/cell/spufs/context.c b/arch/powerpc/platforms/cell/spufs/context.c index 5d6195fc107d..0d88a1c24f67 100644 --- a/arch/powerpc/platforms/cell/spufs/context.c +++ b/arch/powerpc/platforms/cell/spufs/context.c | |||
@@ -45,6 +45,7 @@ struct spu_context *alloc_spu_context(struct address_space *local_store) | |||
45 | init_rwsem(&ctx->state_sema); | 45 | init_rwsem(&ctx->state_sema); |
46 | init_waitqueue_head(&ctx->ibox_wq); | 46 | init_waitqueue_head(&ctx->ibox_wq); |
47 | init_waitqueue_head(&ctx->wbox_wq); | 47 | init_waitqueue_head(&ctx->wbox_wq); |
48 | init_waitqueue_head(&ctx->stop_wq); | ||
48 | ctx->ibox_fasync = NULL; | 49 | ctx->ibox_fasync = NULL; |
49 | ctx->wbox_fasync = NULL; | 50 | ctx->wbox_fasync = NULL; |
50 | ctx->state = SPU_STATE_SAVED; | 51 | ctx->state = SPU_STATE_SAVED; |
@@ -105,7 +106,7 @@ void spu_release(struct spu_context *ctx) | |||
105 | up_read(&ctx->state_sema); | 106 | up_read(&ctx->state_sema); |
106 | } | 107 | } |
107 | 108 | ||
108 | static void spu_unmap_mappings(struct spu_context *ctx) | 109 | void spu_unmap_mappings(struct spu_context *ctx) |
109 | { | 110 | { |
110 | unmap_mapping_range(ctx->local_store, 0, LS_SIZE, 1); | 111 | unmap_mapping_range(ctx->local_store, 0, LS_SIZE, 1); |
111 | } | 112 | } |
@@ -126,7 +127,6 @@ int spu_acquire_runnable(struct spu_context *ctx) | |||
126 | 127 | ||
127 | down_write(&ctx->state_sema); | 128 | down_write(&ctx->state_sema); |
128 | if (ctx->state == SPU_STATE_SAVED) { | 129 | if (ctx->state == SPU_STATE_SAVED) { |
129 | spu_unmap_mappings(ctx); | ||
130 | ret = spu_activate(ctx, 0); | 130 | ret = spu_activate(ctx, 0); |
131 | ctx->state = SPU_STATE_RUNNABLE; | 131 | ctx->state = SPU_STATE_RUNNABLE; |
132 | } | 132 | } |
@@ -154,7 +154,6 @@ void spu_acquire_saved(struct spu_context *ctx) | |||
154 | down_write(&ctx->state_sema); | 154 | down_write(&ctx->state_sema); |
155 | 155 | ||
156 | if (ctx->state == SPU_STATE_RUNNABLE) { | 156 | if (ctx->state == SPU_STATE_RUNNABLE) { |
157 | spu_unmap_mappings(ctx); | ||
158 | spu_deactivate(ctx); | 157 | spu_deactivate(ctx); |
159 | ctx->state = SPU_STATE_SAVED; | 158 | ctx->state = SPU_STATE_SAVED; |
160 | } | 159 | } |
diff --git a/arch/powerpc/platforms/cell/spufs/file.c b/arch/powerpc/platforms/cell/spufs/file.c index 0fe1feccc02d..af5adc372224 100644 --- a/arch/powerpc/platforms/cell/spufs/file.c +++ b/arch/powerpc/platforms/cell/spufs/file.c | |||
@@ -25,6 +25,7 @@ | |||
25 | #include <linux/module.h> | 25 | #include <linux/module.h> |
26 | #include <linux/pagemap.h> | 26 | #include <linux/pagemap.h> |
27 | #include <linux/poll.h> | 27 | #include <linux/poll.h> |
28 | #include <linux/ptrace.h> | ||
28 | 29 | ||
29 | #include <asm/io.h> | 30 | #include <asm/io.h> |
30 | #include <asm/semaphore.h> | 31 | #include <asm/semaphore.h> |
@@ -540,26 +541,122 @@ static struct file_operations spufs_wbox_stat_fops = { | |||
540 | .read = spufs_wbox_stat_read, | 541 | .read = spufs_wbox_stat_read, |
541 | }; | 542 | }; |
542 | 543 | ||
543 | long spufs_run_spu(struct file *file, struct spu_context *ctx, | 544 | /* interrupt-level stop callback function. */ |
544 | u32 *npc, u32 *status) | 545 | void spufs_stop_callback(struct spu *spu) |
546 | { | ||
547 | struct spu_context *ctx = spu->ctx; | ||
548 | |||
549 | wake_up_all(&ctx->stop_wq); | ||
550 | } | ||
551 | |||
552 | static inline int spu_stopped(struct spu_context *ctx, u32 * stat) | ||
553 | { | ||
554 | struct spu *spu; | ||
555 | u64 pte_fault; | ||
556 | |||
557 | *stat = ctx->ops->status_read(ctx); | ||
558 | if (ctx->state != SPU_STATE_RUNNABLE) | ||
559 | return 1; | ||
560 | spu = ctx->spu; | ||
561 | pte_fault = spu->dsisr & | ||
562 | (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED); | ||
563 | return (!(*stat & 0x1) || pte_fault || spu->class_0_pending) ? 1 : 0; | ||
564 | } | ||
565 | |||
566 | static inline int spu_run_init(struct spu_context *ctx, u32 * npc, | ||
567 | u32 * status) | ||
545 | { | 568 | { |
546 | int ret; | 569 | int ret; |
547 | 570 | ||
548 | ret = spu_acquire_runnable(ctx); | 571 | if ((ret = spu_acquire_runnable(ctx)) != 0) |
549 | if (ret) | ||
550 | return ret; | 572 | return ret; |
551 | |||
552 | ctx->ops->npc_write(ctx, *npc); | 573 | ctx->ops->npc_write(ctx, *npc); |
574 | ctx->ops->runcntl_write(ctx, SPU_RUNCNTL_RUNNABLE); | ||
575 | return 0; | ||
576 | } | ||
553 | 577 | ||
554 | ret = spu_run(ctx->spu); | 578 | static inline int spu_run_fini(struct spu_context *ctx, u32 * npc, |
555 | 579 | u32 * status) | |
556 | if (!ret) | 580 | { |
557 | ret = ctx->ops->status_read(ctx); | 581 | int ret = 0; |
558 | 582 | ||
583 | *status = ctx->ops->status_read(ctx); | ||
559 | *npc = ctx->ops->npc_read(ctx); | 584 | *npc = ctx->ops->npc_read(ctx); |
560 | |||
561 | spu_release(ctx); | 585 | spu_release(ctx); |
586 | |||
587 | if (signal_pending(current)) | ||
588 | ret = -ERESTARTSYS; | ||
589 | if (unlikely(current->ptrace & PT_PTRACED)) { | ||
590 | if ((*status & SPU_STATUS_STOPPED_BY_STOP) | ||
591 | && (*status >> SPU_STOP_STATUS_SHIFT) == 0x3fff) { | ||
592 | force_sig(SIGTRAP, current); | ||
593 | ret = -ERESTARTSYS; | ||
594 | } | ||
595 | } | ||
596 | return ret; | ||
597 | } | ||
598 | |||
599 | static inline int spu_reacquire_runnable(struct spu_context *ctx, u32 *npc, | ||
600 | u32 *status) | ||
601 | { | ||
602 | int ret; | ||
603 | |||
604 | if ((ret = spu_run_fini(ctx, npc, status)) != 0) | ||
605 | return ret; | ||
606 | if (*status & (SPU_STATUS_STOPPED_BY_STOP | | ||
607 | SPU_STATUS_STOPPED_BY_HALT)) { | ||
608 | return *status; | ||
609 | } | ||
610 | if ((ret = spu_run_init(ctx, npc, status)) != 0) | ||
611 | return ret; | ||
612 | return 0; | ||
613 | } | ||
614 | |||
615 | static inline int spu_process_events(struct spu_context *ctx) | ||
616 | { | ||
617 | struct spu *spu = ctx->spu; | ||
618 | u64 pte_fault = MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED; | ||
619 | int ret = 0; | ||
620 | |||
621 | if (spu->dsisr & pte_fault) | ||
622 | ret = spu_irq_class_1_bottom(spu); | ||
623 | if (spu->class_0_pending) | ||
624 | ret = spu_irq_class_0_bottom(spu); | ||
625 | if (!ret && signal_pending(current)) | ||
626 | ret = -ERESTARTSYS; | ||
627 | return ret; | ||
628 | } | ||
629 | |||
630 | long spufs_run_spu(struct file *file, struct spu_context *ctx, | ||
631 | u32 * npc, u32 * status) | ||
632 | { | ||
633 | int ret; | ||
634 | |||
635 | if ((ret = spu_run_init(ctx, npc, status)) != 0) | ||
636 | return ret; | ||
637 | |||
638 | do { | ||
639 | ret = spufs_wait(ctx->stop_wq, spu_stopped(ctx, status)); | ||
640 | if (unlikely(ret)) | ||
641 | break; | ||
642 | if (unlikely(ctx->state != SPU_STATE_RUNNABLE)) { | ||
643 | ret = spu_reacquire_runnable(ctx, npc, status); | ||
644 | if (ret) { | ||
645 | return ret; | ||
646 | } | ||
647 | continue; | ||
648 | } | ||
649 | ret = spu_process_events(ctx); | ||
650 | |||
651 | } while (!ret && !(*status & (SPU_STATUS_STOPPED_BY_STOP | | ||
652 | SPU_STATUS_STOPPED_BY_HALT))); | ||
653 | |||
654 | ctx->ops->runcntl_stop(ctx); | ||
655 | ret = spu_run_fini(ctx, npc, status); | ||
656 | if (!ret) | ||
657 | ret = *status; | ||
562 | spu_yield(ctx); | 658 | spu_yield(ctx); |
659 | |||
563 | return ret; | 660 | return ret; |
564 | } | 661 | } |
565 | 662 | ||
diff --git a/arch/powerpc/platforms/cell/spufs/hw_ops.c b/arch/powerpc/platforms/cell/spufs/hw_ops.c index 2e90cae98a87..68812415ee29 100644 --- a/arch/powerpc/platforms/cell/spufs/hw_ops.c +++ b/arch/powerpc/platforms/cell/spufs/hw_ops.c | |||
@@ -186,6 +186,21 @@ static char *spu_hw_get_ls(struct spu_context *ctx) | |||
186 | return ctx->spu->local_store; | 186 | return ctx->spu->local_store; |
187 | } | 187 | } |
188 | 188 | ||
189 | static void spu_hw_runcntl_write(struct spu_context *ctx, u32 val) | ||
190 | { | ||
191 | eieio(); | ||
192 | out_be32(&ctx->spu->problem->spu_runcntl_RW, val); | ||
193 | } | ||
194 | |||
195 | static void spu_hw_runcntl_stop(struct spu_context *ctx) | ||
196 | { | ||
197 | spin_lock_irq(&ctx->spu->register_lock); | ||
198 | out_be32(&ctx->spu->problem->spu_runcntl_RW, SPU_RUNCNTL_STOP); | ||
199 | while (in_be32(&ctx->spu->problem->spu_status_R) & SPU_STATUS_RUNNING) | ||
200 | cpu_relax(); | ||
201 | spin_unlock_irq(&ctx->spu->register_lock); | ||
202 | } | ||
203 | |||
189 | struct spu_context_ops spu_hw_ops = { | 204 | struct spu_context_ops spu_hw_ops = { |
190 | .mbox_read = spu_hw_mbox_read, | 205 | .mbox_read = spu_hw_mbox_read, |
191 | .mbox_stat_read = spu_hw_mbox_stat_read, | 206 | .mbox_stat_read = spu_hw_mbox_stat_read, |
@@ -203,4 +218,6 @@ struct spu_context_ops spu_hw_ops = { | |||
203 | .npc_write = spu_hw_npc_write, | 218 | .npc_write = spu_hw_npc_write, |
204 | .status_read = spu_hw_status_read, | 219 | .status_read = spu_hw_status_read, |
205 | .get_ls = spu_hw_get_ls, | 220 | .get_ls = spu_hw_get_ls, |
221 | .runcntl_write = spu_hw_runcntl_write, | ||
222 | .runcntl_stop = spu_hw_runcntl_stop, | ||
206 | }; | 223 | }; |
diff --git a/arch/powerpc/platforms/cell/spufs/sched.c b/arch/powerpc/platforms/cell/spufs/sched.c index 575027062b0b..e2f10b5b8a6a 100644 --- a/arch/powerpc/platforms/cell/spufs/sched.c +++ b/arch/powerpc/platforms/cell/spufs/sched.c | |||
@@ -119,7 +119,8 @@ static void prio_wakeup(struct spu_runqueue *rq) | |||
119 | } | 119 | } |
120 | } | 120 | } |
121 | 121 | ||
122 | static void prio_wait(struct spu_runqueue *rq, u64 flags) | 122 | static void prio_wait(struct spu_runqueue *rq, struct spu_context *ctx, |
123 | u64 flags) | ||
123 | { | 124 | { |
124 | int prio = current->prio; | 125 | int prio = current->prio; |
125 | wait_queue_head_t *wq = &rq->prio.waitq[prio]; | 126 | wait_queue_head_t *wq = &rq->prio.waitq[prio]; |
@@ -130,9 +131,11 @@ static void prio_wait(struct spu_runqueue *rq, u64 flags) | |||
130 | prepare_to_wait_exclusive(wq, &wait, TASK_INTERRUPTIBLE); | 131 | prepare_to_wait_exclusive(wq, &wait, TASK_INTERRUPTIBLE); |
131 | if (!signal_pending(current)) { | 132 | if (!signal_pending(current)) { |
132 | up(&rq->sem); | 133 | up(&rq->sem); |
134 | up_write(&ctx->state_sema); | ||
133 | pr_debug("%s: pid=%d prio=%d\n", __FUNCTION__, | 135 | pr_debug("%s: pid=%d prio=%d\n", __FUNCTION__, |
134 | current->pid, current->prio); | 136 | current->pid, current->prio); |
135 | schedule(); | 137 | schedule(); |
138 | down_write(&ctx->state_sema); | ||
136 | down(&rq->sem); | 139 | down(&rq->sem); |
137 | } | 140 | } |
138 | finish_wait(wq, &wait); | 141 | finish_wait(wq, &wait); |
@@ -173,7 +176,9 @@ static inline void bind_context(struct spu *spu, struct spu_context *ctx) | |||
173 | mm_needs_global_tlbie(spu->mm); | 176 | mm_needs_global_tlbie(spu->mm); |
174 | spu->ibox_callback = spufs_ibox_callback; | 177 | spu->ibox_callback = spufs_ibox_callback; |
175 | spu->wbox_callback = spufs_wbox_callback; | 178 | spu->wbox_callback = spufs_wbox_callback; |
179 | spu->stop_callback = spufs_stop_callback; | ||
176 | mb(); | 180 | mb(); |
181 | spu_unmap_mappings(ctx); | ||
177 | spu_restore(&ctx->csa, spu); | 182 | spu_restore(&ctx->csa, spu); |
178 | } | 183 | } |
179 | 184 | ||
@@ -181,10 +186,12 @@ static inline void unbind_context(struct spu *spu, struct spu_context *ctx) | |||
181 | { | 186 | { |
182 | pr_debug("%s: unbind pid=%d SPU=%d\n", __FUNCTION__, | 187 | pr_debug("%s: unbind pid=%d SPU=%d\n", __FUNCTION__, |
183 | spu->pid, spu->number); | 188 | spu->pid, spu->number); |
189 | spu_unmap_mappings(ctx); | ||
184 | spu_save(&ctx->csa, spu); | 190 | spu_save(&ctx->csa, spu); |
185 | ctx->state = SPU_STATE_SAVED; | 191 | ctx->state = SPU_STATE_SAVED; |
186 | spu->ibox_callback = NULL; | 192 | spu->ibox_callback = NULL; |
187 | spu->wbox_callback = NULL; | 193 | spu->wbox_callback = NULL; |
194 | spu->stop_callback = NULL; | ||
188 | spu->mm = NULL; | 195 | spu->mm = NULL; |
189 | spu->pid = 0; | 196 | spu->pid = 0; |
190 | spu->prio = MAX_PRIO; | 197 | spu->prio = MAX_PRIO; |
@@ -196,37 +203,35 @@ static inline void unbind_context(struct spu *spu, struct spu_context *ctx) | |||
196 | static struct spu *preempt_active(struct spu_runqueue *rq) | 203 | static struct spu *preempt_active(struct spu_runqueue *rq) |
197 | { | 204 | { |
198 | struct list_head *p; | 205 | struct list_head *p; |
199 | struct spu_context *ctx; | 206 | struct spu *worst, *spu; |
200 | struct spu *spu; | ||
201 | 207 | ||
202 | /* Future: implement real preemption. For now just | 208 | worst = list_entry(rq->active_list.next, struct spu, sched_list); |
203 | * boot a lower priority ctx that is in "detached" | ||
204 | * state, i.e. on a processor but not currently in | ||
205 | * spu_run(). | ||
206 | */ | ||
207 | list_for_each(p, &rq->active_list) { | 209 | list_for_each(p, &rq->active_list) { |
208 | spu = list_entry(p, struct spu, sched_list); | 210 | spu = list_entry(p, struct spu, sched_list); |
209 | if (current->prio < spu->prio) { | 211 | if (spu->prio > worst->prio) { |
210 | ctx = spu->ctx; | 212 | worst = spu; |
211 | if (down_write_trylock(&ctx->state_sema)) { | 213 | } |
212 | if (ctx->state != SPU_STATE_RUNNABLE) { | 214 | } |
213 | up_write(&ctx->state_sema); | 215 | if (current->prio < worst->prio) { |
214 | continue; | 216 | struct spu_context *ctx = worst->ctx; |
215 | } | 217 | |
216 | pr_debug("%s: booting pid=%d from SPU %d\n", | 218 | spu = worst; |
217 | __FUNCTION__, spu->pid, spu->number); | 219 | if (down_write_trylock(&ctx->state_sema)) { |
218 | del_active(rq, spu); | 220 | pr_debug("%s: booting pid=%d from SPU %d\n", |
219 | up(&rq->sem); | 221 | __FUNCTION__, spu->pid, spu->number); |
220 | unbind_context(spu, ctx); | 222 | del_active(rq, spu); |
221 | up_write(&ctx->state_sema); | 223 | up(&rq->sem); |
222 | return spu; | 224 | wake_up_all(&ctx->stop_wq); |
223 | } | 225 | ctx->ops->runcntl_stop(ctx); |
226 | unbind_context(spu, ctx); | ||
227 | up_write(&ctx->state_sema); | ||
228 | return spu; | ||
224 | } | 229 | } |
225 | } | 230 | } |
226 | return NULL; | 231 | return NULL; |
227 | } | 232 | } |
228 | 233 | ||
229 | static struct spu *get_idle_spu(u64 flags) | 234 | static struct spu *get_idle_spu(struct spu_context *ctx, u64 flags) |
230 | { | 235 | { |
231 | struct spu_runqueue *rq; | 236 | struct spu_runqueue *rq; |
232 | struct spu *spu = NULL; | 237 | struct spu *spu = NULL; |
@@ -255,7 +260,7 @@ static struct spu *get_idle_spu(u64 flags) | |||
255 | if ((spu = preempt_active(rq)) != NULL) | 260 | if ((spu = preempt_active(rq)) != NULL) |
256 | return spu; | 261 | return spu; |
257 | } | 262 | } |
258 | prio_wait(rq, flags); | 263 | prio_wait(rq, ctx, flags); |
259 | if (signal_pending(current)) { | 264 | if (signal_pending(current)) { |
260 | prio_wakeup(rq); | 265 | prio_wakeup(rq); |
261 | spu = NULL; | 266 | spu = NULL; |
@@ -322,7 +327,7 @@ int spu_activate(struct spu_context *ctx, u64 flags) | |||
322 | 327 | ||
323 | if (ctx->spu) | 328 | if (ctx->spu) |
324 | return 0; | 329 | return 0; |
325 | spu = get_idle_spu(flags); | 330 | spu = get_idle_spu(ctx, flags); |
326 | if (!spu) | 331 | if (!spu) |
327 | return (signal_pending(current)) ? -ERESTARTSYS : -EAGAIN; | 332 | return (signal_pending(current)) ? -ERESTARTSYS : -EAGAIN; |
328 | bind_context(spu, ctx); | 333 | bind_context(spu, ctx); |
@@ -347,17 +352,19 @@ void spu_deactivate(struct spu_context *ctx) | |||
347 | void spu_yield(struct spu_context *ctx) | 352 | void spu_yield(struct spu_context *ctx) |
348 | { | 353 | { |
349 | struct spu *spu; | 354 | struct spu *spu; |
355 | int need_yield = 0; | ||
350 | 356 | ||
351 | if (!down_write_trylock(&ctx->state_sema)) | 357 | down_write(&ctx->state_sema); |
352 | return; | ||
353 | spu = ctx->spu; | 358 | spu = ctx->spu; |
354 | if ((ctx->state == SPU_STATE_RUNNABLE) && | 359 | if (spu && (sched_find_first_bit(spu->rq->prio.bitmap) < MAX_PRIO)) { |
355 | (sched_find_first_bit(spu->rq->prio.bitmap) <= current->prio)) { | ||
356 | pr_debug("%s: yielding SPU %d\n", __FUNCTION__, spu->number); | 360 | pr_debug("%s: yielding SPU %d\n", __FUNCTION__, spu->number); |
357 | spu_deactivate(ctx); | 361 | spu_deactivate(ctx); |
358 | ctx->state = SPU_STATE_SAVED; | 362 | ctx->state = SPU_STATE_SAVED; |
363 | need_yield = 1; | ||
359 | } | 364 | } |
360 | up_write(&ctx->state_sema); | 365 | up_write(&ctx->state_sema); |
366 | if (unlikely(need_yield)) | ||
367 | yield(); | ||
361 | } | 368 | } |
362 | 369 | ||
363 | int __init spu_sched_init(void) | 370 | int __init spu_sched_init(void) |
diff --git a/arch/powerpc/platforms/cell/spufs/spufs.h b/arch/powerpc/platforms/cell/spufs/spufs.h index 93c6a0537562..20f4e51d1069 100644 --- a/arch/powerpc/platforms/cell/spufs/spufs.h +++ b/arch/powerpc/platforms/cell/spufs/spufs.h | |||
@@ -51,6 +51,7 @@ struct spu_context { | |||
51 | struct kref kref; | 51 | struct kref kref; |
52 | wait_queue_head_t ibox_wq; | 52 | wait_queue_head_t ibox_wq; |
53 | wait_queue_head_t wbox_wq; | 53 | wait_queue_head_t wbox_wq; |
54 | wait_queue_head_t stop_wq; | ||
54 | struct fasync_struct *ibox_fasync; | 55 | struct fasync_struct *ibox_fasync; |
55 | struct fasync_struct *wbox_fasync; | 56 | struct fasync_struct *wbox_fasync; |
56 | struct spu_context_ops *ops; | 57 | struct spu_context_ops *ops; |
@@ -74,6 +75,8 @@ struct spu_context_ops { | |||
74 | void (*npc_write) (struct spu_context * ctx, u32 data); | 75 | void (*npc_write) (struct spu_context * ctx, u32 data); |
75 | u32(*status_read) (struct spu_context * ctx); | 76 | u32(*status_read) (struct spu_context * ctx); |
76 | char*(*get_ls) (struct spu_context * ctx); | 77 | char*(*get_ls) (struct spu_context * ctx); |
78 | void (*runcntl_write) (struct spu_context * ctx, u32 data); | ||
79 | void (*runcntl_stop) (struct spu_context * ctx); | ||
77 | }; | 80 | }; |
78 | 81 | ||
79 | extern struct spu_context_ops spu_hw_ops; | 82 | extern struct spu_context_ops spu_hw_ops; |
@@ -99,6 +102,7 @@ struct spu_context * alloc_spu_context(struct address_space *local_store); | |||
99 | void destroy_spu_context(struct kref *kref); | 102 | void destroy_spu_context(struct kref *kref); |
100 | struct spu_context * get_spu_context(struct spu_context *ctx); | 103 | struct spu_context * get_spu_context(struct spu_context *ctx); |
101 | int put_spu_context(struct spu_context *ctx); | 104 | int put_spu_context(struct spu_context *ctx); |
105 | void spu_unmap_mappings(struct spu_context *ctx); | ||
102 | 106 | ||
103 | void spu_forget(struct spu_context *ctx); | 107 | void spu_forget(struct spu_context *ctx); |
104 | void spu_acquire(struct spu_context *ctx); | 108 | void spu_acquire(struct spu_context *ctx); |
@@ -118,5 +122,6 @@ size_t spu_ibox_read(struct spu_context *ctx, u32 *data); | |||
118 | /* irq callback funcs. */ | 122 | /* irq callback funcs. */ |
119 | void spufs_ibox_callback(struct spu *spu); | 123 | void spufs_ibox_callback(struct spu *spu); |
120 | void spufs_wbox_callback(struct spu *spu); | 124 | void spufs_wbox_callback(struct spu *spu); |
125 | void spufs_stop_callback(struct spu *spu); | ||
121 | 126 | ||
122 | #endif | 127 | #endif |