target/linux/brcm2708/patches-4.19/950-0573-drm-v3d-Add-more-tracepoints-for-V3D-GPU-rendering.patch

   1 From 5d505c19fc3fce8c17711a53287081e61418a776 Mon Sep 17 00:00:00 2001
   2 From: Eric Anholt <eric@anholt.net>
   3 Date: Fri, 30 Nov 2018 16:57:58 -0800
   4 Subject: [PATCH 573/703] drm/v3d: Add more tracepoints for V3D GPU rendering.
   5
   6 The core scheduler tells us when the job is pushed to the scheduler's
   7 queue, and I had the job_run functions saying when they actually queue
   8 the job to the hardware.  By adding tracepoints for the very top of
   9 the ioctls and the IRQs signaling job completion, "perf record -a -e
  10 v3d:.\* -e gpu_scheduler:.\* <job>; perf script" gets you a pretty
  11 decent timeline.
  12
  13 Signed-off-by: Eric Anholt <eric@anholt.net>
  14 Link: https://patchwork.freedesktop.org/patch/msgid/20181201005759.28093-5-eric@anholt.net
  15 Reviewed-by: Dave Emett <david.emett@broadcom.com>
  16 (cherry picked from commit 55a9b74846ed5e6219c7d81a8e1bf96f25d8ad5e)
  17 ---
  18  drivers/gpu/drm/v3d/v3d_gem.c   |   4 ++
  19  drivers/gpu/drm/v3d/v3d_irq.c   |  19 +++++-
  20  drivers/gpu/drm/v3d/v3d_trace.h | 101 ++++++++++++++++++++++++++++++++
  21  3 files changed, 121 insertions(+), 3 deletions(-)
  22
  23 --- a/drivers/gpu/drm/v3d/v3d_gem.c
  24 +++ b/drivers/gpu/drm/v3d/v3d_gem.c
  25 @@ -521,6 +521,8 @@ v3d_submit_cl_ioctl(struct drm_device *d
  26         struct drm_syncobj *sync_out;
  27         int ret = 0;
  28
  29 +       trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
  30 +
  31         if (args->pad != 0) {
  32                 DRM_INFO("pad must be zero: %d\n", args->pad);
  33                 return -EINVAL;
  34 @@ -648,6 +650,8 @@ v3d_submit_tfu_ioctl(struct drm_device *
  35         int ret = 0;
  36         int bo_count;
  37
  38 +       trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia);
  39 +
  40         job = kcalloc(1, sizeof(*job), GFP_KERNEL);
  41         if (!job)
  42                 return -ENOMEM;
  43 --- a/drivers/gpu/drm/v3d/v3d_irq.c
  44 +++ b/drivers/gpu/drm/v3d/v3d_irq.c
  45 @@ -15,6 +15,7 @@
  46
  47  #include "v3d_drv.h"
  48  #include "v3d_regs.h"
  49 +#include "v3d_trace.h"
  50
  51  #define V3D_CORE_IRQS ((u32)(V3D_INT_OUTOMEM | \
  52                              V3D_INT_FLDONE |   \
  53 @@ -88,12 +89,20 @@ v3d_irq(int irq, void *arg)
  54         }
  55
  56         if (intsts & V3D_INT_FLDONE) {
  57 -               dma_fence_signal(v3d->bin_job->bin.done_fence);
  58 +               struct v3d_fence *fence =
  59 +                       to_v3d_fence(v3d->bin_job->bin.done_fence);
  60 +
  61 +               trace_v3d_bcl_irq(&v3d->drm, fence->seqno);
  62 +               dma_fence_signal(&fence->base);
  63                 status = IRQ_HANDLED;
  64         }
  65
  66         if (intsts & V3D_INT_FRDONE) {
  67 -               dma_fence_signal(v3d->render_job->render.done_fence);
  68 +               struct v3d_fence *fence =
  69 +                       to_v3d_fence(v3d->render_job->render.done_fence);
  70 +
  71 +               trace_v3d_rcl_irq(&v3d->drm, fence->seqno);
  72 +               dma_fence_signal(&fence->base);
  73                 status = IRQ_HANDLED;
  74         }
  75
  76 @@ -119,7 +128,11 @@ v3d_hub_irq(int irq, void *arg)
  77         V3D_WRITE(V3D_HUB_INT_CLR, intsts);
  78
  79         if (intsts & V3D_HUB_INT_TFUC) {
  80 -               dma_fence_signal(v3d->tfu_job->done_fence);
  81 +               struct v3d_fence *fence =
  82 +                       to_v3d_fence(v3d->tfu_job->done_fence);
  83 +
  84 +               trace_v3d_tfu_irq(&v3d->drm, fence->seqno);
  85 +               dma_fence_signal(&fence->base);
  86                 status = IRQ_HANDLED;
  87         }
  88
  89 --- a/drivers/gpu/drm/v3d/v3d_trace.h
  90 +++ b/drivers/gpu/drm/v3d/v3d_trace.h
  91 @@ -12,6 +12,28 @@
  92  #define TRACE_SYSTEM v3d
  93  #define TRACE_INCLUDE_FILE v3d_trace
  94
  95 +TRACE_EVENT(v3d_submit_cl_ioctl,
  96 +           TP_PROTO(struct drm_device *dev, u32 ct1qba, u32 ct1qea),
  97 +           TP_ARGS(dev, ct1qba, ct1qea),
  98 +
  99 +           TP_STRUCT__entry(
 100 +                            __field(u32, dev)
 101 +                            __field(u32, ct1qba)
 102 +                            __field(u32, ct1qea)
 103 +                            ),
 104 +
 105 +           TP_fast_assign(
 106 +                          __entry->dev = dev->primary->index;
 107 +                          __entry->ct1qba = ct1qba;
 108 +                          __entry->ct1qea = ct1qea;
 109 +                          ),
 110 +
 111 +           TP_printk("dev=%u, RCL 0x%08x..0x%08x",
 112 +                     __entry->dev,
 113 +                     __entry->ct1qba,
 114 +                     __entry->ct1qea)
 115 +);
 116 +
 117  TRACE_EVENT(v3d_submit_cl,
 118             TP_PROTO(struct drm_device *dev, bool is_render,
 119                      uint64_t seqno,
 120 @@ -42,6 +64,85 @@ TRACE_EVENT(v3d_submit_cl,
 121                       __entry->ctnqea)
 122  );
 123
 124 +TRACE_EVENT(v3d_bcl_irq,
 125 +           TP_PROTO(struct drm_device *dev,
 126 +                    uint64_t seqno),
 127 +           TP_ARGS(dev, seqno),
 128 +
 129 +           TP_STRUCT__entry(
 130 +                            __field(u32, dev)
 131 +                            __field(u64, seqno)
 132 +                            ),
 133 +
 134 +           TP_fast_assign(
 135 +                          __entry->dev = dev->primary->index;
 136 +                          __entry->seqno = seqno;
 137 +                          ),
 138 +
 139 +           TP_printk("dev=%u, seqno=%llu",
 140 +                     __entry->dev,
 141 +                     __entry->seqno)
 142 +);
 143 +
 144 +TRACE_EVENT(v3d_rcl_irq,
 145 +           TP_PROTO(struct drm_device *dev,
 146 +                    uint64_t seqno),
 147 +           TP_ARGS(dev, seqno),
 148 +
 149 +           TP_STRUCT__entry(
 150 +                            __field(u32, dev)
 151 +                            __field(u64, seqno)
 152 +                            ),
 153 +
 154 +           TP_fast_assign(
 155 +                          __entry->dev = dev->primary->index;
 156 +                          __entry->seqno = seqno;
 157 +                          ),
 158 +
 159 +           TP_printk("dev=%u, seqno=%llu",
 160 +                     __entry->dev,
 161 +                     __entry->seqno)
 162 +);
 163 +
 164 +TRACE_EVENT(v3d_tfu_irq,
 165 +           TP_PROTO(struct drm_device *dev,
 166 +                    uint64_t seqno),
 167 +           TP_ARGS(dev, seqno),
 168 +
 169 +           TP_STRUCT__entry(
 170 +                            __field(u32, dev)
 171 +                            __field(u64, seqno)
 172 +                            ),
 173 +
 174 +           TP_fast_assign(
 175 +                          __entry->dev = dev->primary->index;
 176 +                          __entry->seqno = seqno;
 177 +                          ),
 178 +
 179 +           TP_printk("dev=%u, seqno=%llu",
 180 +                     __entry->dev,
 181 +                     __entry->seqno)
 182 +);
 183 +
 184 +TRACE_EVENT(v3d_submit_tfu_ioctl,
 185 +           TP_PROTO(struct drm_device *dev, u32 iia),
 186 +           TP_ARGS(dev, iia),
 187 +
 188 +           TP_STRUCT__entry(
 189 +                            __field(u32, dev)
 190 +                            __field(u32, iia)
 191 +                            ),
 192 +
 193 +           TP_fast_assign(
 194 +                          __entry->dev = dev->primary->index;
 195 +                          __entry->iia = iia;
 196 +                          ),
 197 +
 198 +           TP_printk("dev=%u, IIA 0x%08x",
 199 +                     __entry->dev,
 200 +                     __entry->iia)
 201 +);
 202 +
 203  TRACE_EVENT(v3d_submit_tfu,
 204             TP_PROTO(struct drm_device *dev,
 205                      uint64_t seqno),