gpu: nvgpu: support tuning per-ch deterministic opts

Add a new ioctl NVGPU_GPU_IOCTL_SET_DETERMINISTIC_OPTS to adjust
deterministic options on a per-channel basis. Currently, the only
supported option is to relax the no-railgating requirement on open
deterministic channels. This also disallows submits on such channels,
until the railgate option is reset.

Bug 200327089

Change-Id: If4f0f51fd1d40ad7407d13638150d7402479aff0
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1554563
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

6 files changed, 137 insertions, 1 deletions

diff --git a/drivers/gpu/nvgpu/common/linux/channel.c b/drivers/gpu/nvgpu/common/linux/channel.c
index 716c5820..c295336f 100644
--- a/drivers/gpu/nvgpu/common/linux/channel.c
+++ b/drivers/gpu/nvgpu/common/linux/channel.c
@@ -537,6 +537,17 @@ int gk20a_submit_channel_gpfifo(struct channel_gk20a *c,
         if (c->deterministic)
                 nvgpu_rwsem_down_read(&g->deterministic_busy);
 
+        if (c->deterministic && c->deterministic_railgate_allowed) {
+                /*
+                 * Nope - this channel has dropped its own power ref. As
+                 * deterministic submits don't hold power on per each submitted
+                 * job like normal ones do, the GPU might railgate any time now
+                 * and thus submit is disallowed.
+                 */
+                err = -EINVAL;
+                goto clean_up;
+        }
+
         trace_gk20a_channel_submit_gpfifo(g->name,
                                           c->chid,
                                           num_entries,
diff --git a/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c b/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c
index 73911717..08a831b9 100644
--- a/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c
+++ b/drivers/gpu/nvgpu/common/linux/ioctl_ctrl.c
@@ -161,6 +161,8 @@ static struct nvgpu_flags_mapping flags_mapping[] = {
                 NVGPU_SUPPORT_DETERMINISTIC_SUBMIT_NO_JOBTRACKING},
         {NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_FULL,
                 NVGPU_SUPPORT_DETERMINISTIC_SUBMIT_FULL},
+        {NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_OPTS,
+                NVGPU_SUPPORT_DETERMINISTIC_OPTS},
         {NVGPU_GPU_FLAGS_SUPPORT_IO_COHERENCE,
                 NVGPU_SUPPORT_IO_COHERENCE},
         {NVGPU_GPU_FLAGS_SUPPORT_RESCHEDULE_RUNLIST,
@@ -1319,6 +1321,114 @@ static int nvgpu_gpu_set_therm_alert_limit(struct gk20a *g,
         return err;
 }
 
+static int nvgpu_gpu_set_deterministic_ch_railgate(struct channel_gk20a *ch,
+                u32 flags)
+{
+        int err = 0;
+        bool allow;
+        bool disallow;
+
+        allow = flags &
+                NVGPU_GPU_SET_DETERMINISTIC_OPTS_FLAGS_ALLOW_RAILGATING;
+
+        disallow = flags &
+                NVGPU_GPU_SET_DETERMINISTIC_OPTS_FLAGS_DISALLOW_RAILGATING;
+
+        /* Can't be both at the same time */
+        if (allow && disallow)
+                return -EINVAL;
+
+        /* Nothing to do */
+        if (!allow && !disallow)
+                return 0;
+
+        /*
+         * Moving into explicit idle or back from it? A call that doesn't
+         * change the status is a no-op.
+         */
+        if (!ch->deterministic_railgate_allowed &&
+                        allow) {
+                gk20a_idle(ch->g);
+        } else if (ch->deterministic_railgate_allowed &&
+                        !allow) {
+                err = gk20a_busy(ch->g);
+                if (err) {
+                        nvgpu_warn(ch->g,
+                                "cannot busy to restore deterministic ch");
+                        return err;
+                }
+        }
+        ch->deterministic_railgate_allowed = allow;
+
+        return err;
+}
+
+static int nvgpu_gpu_set_deterministic_ch(struct channel_gk20a *ch, u32 flags)
+{
+        if (!ch->deterministic)
+                return -EINVAL;
+
+        return nvgpu_gpu_set_deterministic_ch_railgate(ch, flags);
+}
+
+static int nvgpu_gpu_set_deterministic_opts(struct gk20a *g,
+                struct nvgpu_gpu_set_deterministic_opts_args *args)
+{
+        int __user *user_channels;
+        u32 i = 0;
+        int err = 0;
+
+        gk20a_dbg_fn("");
+
+        user_channels = (int __user *)(uintptr_t)args->channels;
+
+        /* Upper limit; prevent holding deterministic_busy for long */
+        if (args->num_channels > g->fifo.num_channels) {
+                err = -EINVAL;
+                goto out;
+        }
+
+        /* Trivial sanity check first */
+        if (!access_ok(VERIFY_READ, user_channels,
+                                args->num_channels * sizeof(int))) {
+                err = -EFAULT;
+                goto out;
+        }
+
+        nvgpu_rwsem_down_read(&g->deterministic_busy);
+
+        /* note: we exit at the first failure */
+        for (; i < args->num_channels; i++) {
+                int ch_fd = 0;
+                struct channel_gk20a *ch;
+
+                if (copy_from_user(&ch_fd, &user_channels[i], sizeof(int))) {
+                        /* User raced with above access_ok */
+                        err = -EFAULT;
+                        break;
+                }
+
+                ch = gk20a_get_channel_from_file(ch_fd);
+                if (!ch) {
+                        err = -EINVAL;
+                        break;
+                }
+
+                err = nvgpu_gpu_set_deterministic_ch(ch, args->flags);
+
+                gk20a_channel_put(ch);
+
+                if (err)
+                        break;
+        }
+
+        nvgpu_rwsem_up_read(&g->deterministic_busy);
+
+out:
+        args->num_channels = i;
+        return err;
+}
+
 long gk20a_ctrl_dev_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 {
         struct gk20a_ctrl_priv *priv = filp->private_data;
@@ -1633,6 +1743,11 @@ long gk20a_ctrl_dev_ioctl(struct file *filp, unsigned int cmd, unsigned long arg
                         (struct nvgpu_gpu_set_therm_alert_limit_args *)buf);
                 break;
 
+        case NVGPU_GPU_IOCTL_SET_DETERMINISTIC_OPTS:
+                err = nvgpu_gpu_set_deterministic_opts(g,
+                        (struct nvgpu_gpu_set_deterministic_opts_args *)buf);
+                break;
+
         default:
                 gk20a_dbg_info("unrecognized gpu ioctl cmd: 0x%x", cmd);
                 err = -ENOTTY;
diff --git a/drivers/gpu/nvgpu/gk20a/channel_gk20a.c b/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
index 805902eb..6c607ae2 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
@@ -578,7 +578,10 @@ unbind:
         if (ch->deterministic) {
                 nvgpu_rwsem_down_read(&g->deterministic_busy);
                 ch->deterministic = false;
-                gk20a_idle(g);
+                if (!ch->deterministic_railgate_allowed)
+                        gk20a_idle(g);
+                ch->deterministic_railgate_allowed = false;
+
                 nvgpu_rwsem_up_read(&g->deterministic_busy);
         }
 
diff --git a/drivers/gpu/nvgpu/gk20a/channel_gk20a.h b/drivers/gpu/nvgpu/gk20a/channel_gk20a.h
index 762daea9..0cb60200 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/channel_gk20a.h
@@ -208,6 +208,8 @@ struct channel_gk20a {
         bool first_init;
         bool vpr;
         bool deterministic;
+        /* deterministic, but explicitly idle and submits disallowed */
+        bool deterministic_railgate_allowed;
         bool cde;
         pid_t pid;
         pid_t tgid;
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.c b/drivers/gpu/nvgpu/gk20a/gk20a.c
index e8db9d2c..756db415 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.c
@@ -439,6 +439,8 @@ int gk20a_init_gpu_characteristics(struct gk20a *g)
                                 NVGPU_SUPPORT_DETERMINISTIC_SUBMIT_FULL,
                                 true);
 
+        __nvgpu_set_enabled(g, NVGPU_SUPPORT_DETERMINISTIC_OPTS, true);
+
         __nvgpu_set_enabled(g, NVGPU_SUPPORT_USERSPACE_MANAGED_AS, true);
         __nvgpu_set_enabled(g, NVGPU_SUPPORT_TSG, true);
 
diff --git a/drivers/gpu/nvgpu/include/nvgpu/enabled.h b/drivers/gpu/nvgpu/include/nvgpu/enabled.h
index 14595497..82f9e6d2 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/enabled.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/enabled.h
@@ -91,6 +91,9 @@ struct gk20a;
 /* FECS context switch tracing is available */
 #define NVGPU_SUPPORT_FECS_CTXSW_TRACE          38
 
+/* NVGPU_GPU_IOCTL_SET_DETERMINISTIC_OPTS is available */
+#define NVGPU_SUPPORT_DETERMINISTIC_OPTS        39
+
 /*
  * Security flags
  */