diff options
Diffstat (limited to 'drivers')
-rw-r--r-- | drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c | 78 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/bus_gk20a.c | 61 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/bus_gk20a.h | 9 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gk20a/gk20a.h | 4 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/gm20b/bus_gm20b.c | 1 | ||||
-rw-r--r-- | drivers/gpu/nvgpu/vgpu/vgpu.c | 43 |
6 files changed, 136 insertions, 60 deletions
diff --git a/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c b/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c index 90a3fbd4..025a30fe 100644 --- a/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c +++ b/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c | |||
@@ -533,76 +533,34 @@ static int gk20a_ctrl_get_buffer_info( | |||
533 | &args->out.id, &args->out.length); | 533 | &args->out.id, &args->out.length); |
534 | } | 534 | } |
535 | 535 | ||
536 | static inline u64 get_cpu_timestamp_tsc(void) | ||
537 | { | ||
538 | return ((u64) get_cycles()); | ||
539 | } | ||
540 | |||
541 | static inline u64 get_cpu_timestamp_jiffies(void) | ||
542 | { | ||
543 | return (get_jiffies_64() - INITIAL_JIFFIES); | ||
544 | } | ||
545 | |||
546 | static inline u64 get_cpu_timestamp_timeofday(void) | ||
547 | { | ||
548 | struct timeval tv; | ||
549 | |||
550 | do_gettimeofday(&tv); | ||
551 | return timeval_to_jiffies(&tv); | ||
552 | } | ||
553 | |||
554 | static inline int get_timestamps_zipper(struct gk20a *g, | ||
555 | u64 (*get_cpu_timestamp)(void), | ||
556 | struct nvgpu_gpu_get_cpu_time_correlation_info_args *args) | ||
557 | { | ||
558 | int err = 0; | ||
559 | unsigned int i = 0; | ||
560 | |||
561 | if (gk20a_busy(g)) { | ||
562 | nvgpu_err(g, "GPU not powered on"); | ||
563 | err = -EINVAL; | ||
564 | goto end; | ||
565 | } | ||
566 | |||
567 | for (i = 0; i < args->count; i++) { | ||
568 | err = g->ops.bus.read_ptimer(g, &args->samples[i].gpu_timestamp); | ||
569 | if (err) | ||
570 | return err; | ||
571 | |||
572 | args->samples[i].cpu_timestamp = get_cpu_timestamp(); | ||
573 | } | ||
574 | |||
575 | end: | ||
576 | gk20a_idle(g); | ||
577 | return err; | ||
578 | } | ||
579 | |||
580 | static int nvgpu_gpu_get_cpu_time_correlation_info( | 536 | static int nvgpu_gpu_get_cpu_time_correlation_info( |
581 | struct gk20a *g, | 537 | struct gk20a *g, |
582 | struct nvgpu_gpu_get_cpu_time_correlation_info_args *args) | 538 | struct nvgpu_gpu_get_cpu_time_correlation_info_args *args) |
583 | { | 539 | { |
584 | int err = 0; | 540 | struct nvgpu_cpu_time_correlation_sample *samples; |
585 | u64 (*get_cpu_timestamp)(void) = NULL; | 541 | int err; |
542 | u32 i; | ||
586 | 543 | ||
587 | if (args->count > NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_MAX_COUNT) | 544 | if (args->count > NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_MAX_COUNT) |
588 | return -EINVAL; | 545 | return -EINVAL; |
589 | 546 | ||
590 | switch (args->source_id) { | 547 | samples = nvgpu_kzalloc(g, args->count * |
591 | case NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_SRC_ID_TSC: | 548 | sizeof(struct nvgpu_cpu_time_correlation_sample)); |
592 | get_cpu_timestamp = get_cpu_timestamp_tsc; | 549 | if (!samples) { |
593 | break; | 550 | return -ENOMEM; |
594 | case NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_SRC_ID_JIFFIES: | ||
595 | get_cpu_timestamp = get_cpu_timestamp_jiffies; | ||
596 | break; | ||
597 | case NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_SRC_ID_TIMEOFDAY: | ||
598 | get_cpu_timestamp = get_cpu_timestamp_timeofday; | ||
599 | break; | ||
600 | default: | ||
601 | nvgpu_err(g, "invalid cpu clock source id"); | ||
602 | return -EINVAL; | ||
603 | } | 551 | } |
604 | 552 | ||
605 | err = get_timestamps_zipper(g, get_cpu_timestamp, args); | 553 | err = g->ops.bus.get_timestamps_zipper(g, |
554 | args->source_id, args->count, samples); | ||
555 | if (!err) { | ||
556 | for (i = 0; i < args->count; i++) { | ||
557 | args->samples[i].cpu_timestamp = samples[i].cpu_timestamp; | ||
558 | args->samples[i].gpu_timestamp = samples[i].gpu_timestamp; | ||
559 | } | ||
560 | } | ||
561 | |||
562 | nvgpu_kfree(g, samples); | ||
563 | |||
606 | return err; | 564 | return err; |
607 | } | 565 | } |
608 | 566 | ||
diff --git a/drivers/gpu/nvgpu/gk20a/bus_gk20a.c b/drivers/gpu/nvgpu/gk20a/bus_gk20a.c index 52ef08e4..7f0ca013 100644 --- a/drivers/gpu/nvgpu/gk20a/bus_gk20a.c +++ b/drivers/gpu/nvgpu/gk20a/bus_gk20a.c | |||
@@ -128,6 +128,66 @@ int gk20a_read_ptimer(struct gk20a *g, u64 *value) | |||
128 | return -EBUSY; | 128 | return -EBUSY; |
129 | } | 129 | } |
130 | 130 | ||
131 | static inline u64 get_cpu_timestamp_tsc(void) | ||
132 | { | ||
133 | return ((u64) get_cycles()); | ||
134 | } | ||
135 | |||
136 | static inline u64 get_cpu_timestamp_jiffies(void) | ||
137 | { | ||
138 | return (get_jiffies_64() - INITIAL_JIFFIES); | ||
139 | } | ||
140 | |||
141 | static inline u64 get_cpu_timestamp_timeofday(void) | ||
142 | { | ||
143 | struct timeval tv; | ||
144 | |||
145 | do_gettimeofday(&tv); | ||
146 | return timeval_to_jiffies(&tv); | ||
147 | } | ||
148 | |||
149 | int gk20a_get_timestamps_zipper(struct gk20a *g, | ||
150 | u32 source_id, u32 count, | ||
151 | struct nvgpu_cpu_time_correlation_sample *samples) | ||
152 | { | ||
153 | int err = 0; | ||
154 | unsigned int i = 0; | ||
155 | u64 (*get_cpu_timestamp)(void) = NULL; | ||
156 | |||
157 | switch (source_id) { | ||
158 | case NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_SRC_ID_TSC: | ||
159 | get_cpu_timestamp = get_cpu_timestamp_tsc; | ||
160 | break; | ||
161 | case NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_SRC_ID_JIFFIES: | ||
162 | get_cpu_timestamp = get_cpu_timestamp_jiffies; | ||
163 | break; | ||
164 | case NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_SRC_ID_TIMEOFDAY: | ||
165 | get_cpu_timestamp = get_cpu_timestamp_timeofday; | ||
166 | break; | ||
167 | default: | ||
168 | nvgpu_err(g, "invalid cpu clock source id\n"); | ||
169 | return -EINVAL; | ||
170 | } | ||
171 | |||
172 | if (gk20a_busy(g)) { | ||
173 | nvgpu_err(g, "GPU not powered on\n"); | ||
174 | err = -EINVAL; | ||
175 | goto end; | ||
176 | } | ||
177 | |||
178 | for (i = 0; i < count; i++) { | ||
179 | err = g->ops.bus.read_ptimer(g, &samples[i].gpu_timestamp); | ||
180 | if (err) | ||
181 | return err; | ||
182 | |||
183 | samples[i].cpu_timestamp = get_cpu_timestamp(); | ||
184 | } | ||
185 | |||
186 | end: | ||
187 | gk20a_idle(g); | ||
188 | return err; | ||
189 | } | ||
190 | |||
131 | static int gk20a_bus_bar1_bind(struct gk20a *g, struct nvgpu_mem *bar1_inst) | 191 | static int gk20a_bus_bar1_bind(struct gk20a *g, struct nvgpu_mem *bar1_inst) |
132 | { | 192 | { |
133 | u64 iova = gk20a_mm_inst_block_addr(g, bar1_inst); | 193 | u64 iova = gk20a_mm_inst_block_addr(g, bar1_inst); |
@@ -150,5 +210,6 @@ void gk20a_init_bus(struct gpu_ops *gops) | |||
150 | gops->bus.init_hw = gk20a_bus_init_hw; | 210 | gops->bus.init_hw = gk20a_bus_init_hw; |
151 | gops->bus.isr = gk20a_bus_isr; | 211 | gops->bus.isr = gk20a_bus_isr; |
152 | gops->bus.read_ptimer = gk20a_read_ptimer; | 212 | gops->bus.read_ptimer = gk20a_read_ptimer; |
213 | gops->bus.get_timestamps_zipper = gk20a_get_timestamps_zipper; | ||
153 | gops->bus.bar1_bind = gk20a_bus_bar1_bind; | 214 | gops->bus.bar1_bind = gk20a_bus_bar1_bind; |
154 | } | 215 | } |
diff --git a/drivers/gpu/nvgpu/gk20a/bus_gk20a.h b/drivers/gpu/nvgpu/gk20a/bus_gk20a.h index 72710c35..088c385e 100644 --- a/drivers/gpu/nvgpu/gk20a/bus_gk20a.h +++ b/drivers/gpu/nvgpu/gk20a/bus_gk20a.h | |||
@@ -22,10 +22,19 @@ struct gk20a; | |||
22 | struct gpu_ops; | 22 | struct gpu_ops; |
23 | struct nvgpu_mem; | 23 | struct nvgpu_mem; |
24 | 24 | ||
25 | struct nvgpu_cpu_time_correlation_sample { | ||
26 | u64 cpu_timestamp; | ||
27 | u64 gpu_timestamp; | ||
28 | }; | ||
29 | |||
25 | void gk20a_init_bus(struct gpu_ops *gops); | 30 | void gk20a_init_bus(struct gpu_ops *gops); |
26 | 31 | ||
27 | void gk20a_bus_isr(struct gk20a *g); | 32 | void gk20a_bus_isr(struct gk20a *g); |
28 | int gk20a_read_ptimer(struct gk20a *g, u64 *value); | 33 | int gk20a_read_ptimer(struct gk20a *g, u64 *value); |
29 | void gk20a_bus_init_hw(struct gk20a *g); | 34 | void gk20a_bus_init_hw(struct gk20a *g); |
30 | 35 | ||
36 | int gk20a_get_timestamps_zipper(struct gk20a *g, | ||
37 | u32 source_id, u32 count, | ||
38 | struct nvgpu_cpu_time_correlation_sample *samples); | ||
39 | |||
31 | #endif /* GK20A_H */ | 40 | #endif /* GK20A_H */ |
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.h b/drivers/gpu/nvgpu/gk20a/gk20a.h index 34f89fc8..d1a55104 100644 --- a/drivers/gpu/nvgpu/gk20a/gk20a.h +++ b/drivers/gpu/nvgpu/gk20a/gk20a.h | |||
@@ -63,6 +63,7 @@ struct nvgpu_nvhost_dev; | |||
63 | #include "priv_ring_gk20a.h" | 63 | #include "priv_ring_gk20a.h" |
64 | #include "therm_gk20a.h" | 64 | #include "therm_gk20a.h" |
65 | #include "gm20b/acr_gm20b.h" | 65 | #include "gm20b/acr_gm20b.h" |
66 | #include "gk20a/bus_gk20a.h" | ||
66 | #include "cde_gk20a.h" | 67 | #include "cde_gk20a.h" |
67 | #include "sched_gk20a.h" | 68 | #include "sched_gk20a.h" |
68 | #ifdef CONFIG_ARCH_TEGRA_18x_SOC | 69 | #ifdef CONFIG_ARCH_TEGRA_18x_SOC |
@@ -905,6 +906,9 @@ struct gpu_ops { | |||
905 | void (*init_hw)(struct gk20a *g); | 906 | void (*init_hw)(struct gk20a *g); |
906 | void (*isr)(struct gk20a *g); | 907 | void (*isr)(struct gk20a *g); |
907 | int (*read_ptimer)(struct gk20a *g, u64 *value); | 908 | int (*read_ptimer)(struct gk20a *g, u64 *value); |
909 | int (*get_timestamps_zipper)(struct gk20a *g, | ||
910 | u32 source_id, u32 count, | ||
911 | struct nvgpu_cpu_time_correlation_sample *); | ||
908 | int (*bar1_bind)(struct gk20a *g, struct nvgpu_mem *bar1_inst); | 912 | int (*bar1_bind)(struct gk20a *g, struct nvgpu_mem *bar1_inst); |
909 | } bus; | 913 | } bus; |
910 | 914 | ||
diff --git a/drivers/gpu/nvgpu/gm20b/bus_gm20b.c b/drivers/gpu/nvgpu/gm20b/bus_gm20b.c index 7b414196..39778c55 100644 --- a/drivers/gpu/nvgpu/gm20b/bus_gm20b.c +++ b/drivers/gpu/nvgpu/gm20b/bus_gm20b.c | |||
@@ -58,5 +58,6 @@ void gm20b_init_bus(struct gpu_ops *gops) | |||
58 | gops->bus.init_hw = gk20a_bus_init_hw; | 58 | gops->bus.init_hw = gk20a_bus_init_hw; |
59 | gops->bus.isr = gk20a_bus_isr; | 59 | gops->bus.isr = gk20a_bus_isr; |
60 | gops->bus.read_ptimer = gk20a_read_ptimer; | 60 | gops->bus.read_ptimer = gk20a_read_ptimer; |
61 | gops->bus.get_timestamps_zipper = gk20a_get_timestamps_zipper; | ||
61 | gops->bus.bar1_bind = gm20b_bus_bar1_bind; | 62 | gops->bus.bar1_bind = gm20b_bus_bar1_bind; |
62 | } | 63 | } |
diff --git a/drivers/gpu/nvgpu/vgpu/vgpu.c b/drivers/gpu/nvgpu/vgpu/vgpu.c index cdd0d378..a8c28826 100644 --- a/drivers/gpu/nvgpu/vgpu/vgpu.c +++ b/drivers/gpu/nvgpu/vgpu/vgpu.c | |||
@@ -367,6 +367,48 @@ static int vgpu_read_ptimer(struct gk20a *g, u64 *value) | |||
367 | return err; | 367 | return err; |
368 | } | 368 | } |
369 | 369 | ||
370 | int vgpu_get_timestamps_zipper(struct gk20a *g, | ||
371 | u32 source_id, u32 count, | ||
372 | struct nvgpu_cpu_time_correlation_sample *samples) | ||
373 | { | ||
374 | struct tegra_vgpu_cmd_msg msg = {0}; | ||
375 | struct tegra_vgpu_get_timestamps_zipper_params *p = | ||
376 | &msg.params.get_timestamps_zipper; | ||
377 | int err; | ||
378 | u32 i; | ||
379 | |||
380 | gk20a_dbg_fn(""); | ||
381 | |||
382 | if (count > TEGRA_VGPU_GET_TIMESTAMPS_ZIPPER_MAX_COUNT) { | ||
383 | nvgpu_err(g, "count %u overflow", count); | ||
384 | return -EINVAL; | ||
385 | } | ||
386 | |||
387 | if (source_id != NVGPU_GPU_GET_CPU_TIME_CORRELATION_INFO_SRC_ID_TSC) { | ||
388 | nvgpu_err(g, "source_id %u not supported", source_id); | ||
389 | return -EINVAL; | ||
390 | } | ||
391 | |||
392 | msg.cmd = TEGRA_VGPU_CMD_GET_TIMESTAMPS_ZIPPER; | ||
393 | msg.handle = vgpu_get_handle(g); | ||
394 | p->source_id = TEGRA_VGPU_GET_TIMESTAMPS_ZIPPER_SRC_ID_TSC; | ||
395 | p->count = count; | ||
396 | |||
397 | err = vgpu_comm_sendrecv(&msg, sizeof(msg), sizeof(msg)); | ||
398 | err = err ? err : msg.ret; | ||
399 | if (err) { | ||
400 | nvgpu_err(g, "vgpu get timestamps zipper failed, err=%d", err); | ||
401 | return err; | ||
402 | } | ||
403 | |||
404 | for (i = 0; i < count; i++) { | ||
405 | samples[i].cpu_timestamp = p->samples[i].cpu_timestamp; | ||
406 | samples[i].gpu_timestamp = p->samples[i].gpu_timestamp; | ||
407 | } | ||
408 | |||
409 | return err; | ||
410 | } | ||
411 | |||
370 | void vgpu_init_hal_common(struct gk20a *g) | 412 | void vgpu_init_hal_common(struct gk20a *g) |
371 | { | 413 | { |
372 | struct gpu_ops *gops = &g->ops; | 414 | struct gpu_ops *gops = &g->ops; |
@@ -384,6 +426,7 @@ void vgpu_init_hal_common(struct gk20a *g) | |||
384 | #endif | 426 | #endif |
385 | gops->chip_init_gpu_characteristics = vgpu_init_gpu_characteristics; | 427 | gops->chip_init_gpu_characteristics = vgpu_init_gpu_characteristics; |
386 | gops->bus.read_ptimer = vgpu_read_ptimer; | 428 | gops->bus.read_ptimer = vgpu_read_ptimer; |
429 | gops->bus.get_timestamps_zipper = vgpu_get_timestamps_zipper; | ||
387 | } | 430 | } |
388 | 431 | ||
389 | static int vgpu_init_hal(struct gk20a *g) | 432 | static int vgpu_init_hal(struct gk20a *g) |