gpu: nvgpu: add sched domain bind ioctl

Support binding TSGs to some other scheduling domain than the default
one. Binding happens by name until a more robust interface appears in
the future, as the name is a natural identifier for users.

No other domains are actually created anywhere yet; that will happen in
next patches.

Jira NVGPU-6788

Change-Id: I5abcdea869b525b0a0e9937302f106f7eee94ec2
Signed-off-by: Konsta Hölttä <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvgpu/+/2628047
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
Reviewed-by: svc-mobile-cert <svc-mobile-cert@nvidia.com>
Reviewed-by: Alex Waterman <alexw@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
GVS: Gerrit_Virtual_Submit

drivers/gpu/nvgpu/common/fifo/tsg.c

@@ -109,7 +109,13 @@ int nvgpu_tsg_bind_channel(struct nvgpu_tsg *tsg, struct nvgpu_channel *ch)
 	 */
 	if (tsg->runlist == NULL) {
 		tsg->runlist = ch->runlist;
-		tsg->rl_domain = nvgpu_rl_domain_get(g, tsg->runlist->id, "(default)");
+		/*
+		 * The rl domain identifier is stashed in tsg->rl_domain->name
+		 * when the tsg is bound to a domain, but at that point there
+		 * are no channels yet to describe which runlist id should be
+		 * used. Now we know.
+		 */
+		tsg->rl_domain = nvgpu_rl_domain_get(g, tsg->runlist->id, tsg->rl_domain->name);
 		WARN_ON(tsg->rl_domain == NULL);
 	} else {
 		if (tsg->runlist != ch->runlist) {
@@ -147,6 +153,30 @@ int nvgpu_tsg_bind_channel(struct nvgpu_tsg *tsg, struct nvgpu_channel *ch)
 	return err;
 }
 
+int nvgpu_tsg_bind_domain(struct nvgpu_tsg *tsg, const char *domain_name)
+{
+	struct nvgpu_runlist_domain *domain;
+	struct gk20a *g = tsg->g;
+
+	/* Hopping channels from one domain to another is not allowed */
+	if (tsg->num_active_channels != 0U) {
+		return -EINVAL;
+	}
+
+	/*
+	 * The domain ptr will get updated with the right id once the runlist
+	 * gets specified based on the first channel.
+	 */
+	domain = nvgpu_rl_domain_get(g, 0, domain_name);
+	if (domain == NULL) {
+		return -ENOENT;
+	}
+
+	tsg->rl_domain = domain;
+
+	return 0;
+}
+
 static bool nvgpu_tsg_is_multi_channel(struct nvgpu_tsg *tsg)
 {
 	bool ret = false;
@@ -807,6 +837,11 @@ int nvgpu_tsg_open_common(struct gk20a *g, struct nvgpu_tsg *tsg, pid_t pid)
 	tsg->interleave_level = NVGPU_FIFO_RUNLIST_INTERLEAVE_LEVEL_LOW;
 	tsg->timeslice_us = g->ops.tsg.default_timeslice_us(g);
 	tsg->runlist = NULL;
+	/*
+	 * The domain ptr will get updated with the right id once the runlist
+	 * gets specified based on the first channel.
+	 */
+	tsg->rl_domain = nvgpu_rl_domain_get(g, 0, "(default)");
 #ifdef CONFIG_NVGPU_DEBUGGER
 	tsg->sm_exception_mask_type = NVGPU_SM_EXCEPTION_TYPE_MASK_NONE;
 #endif

drivers/gpu/nvgpu/include/nvgpu/tsg.h

@@ -373,6 +373,8 @@ void nvgpu_tsg_disable(struct nvgpu_tsg *tsg);
 int nvgpu_tsg_bind_channel(struct nvgpu_tsg *tsg,
 			struct nvgpu_channel *ch);
 
+int nvgpu_tsg_bind_domain(struct nvgpu_tsg *tsg, const char *domain_name);
+
 /**
  * @brief Get pointer to #nvgpu_tsg for the tsgid.
  *

drivers/gpu/nvgpu/os/linux/ioctl_tsg.c

@@ -176,6 +176,13 @@ out:
 	return err;
 }
 
+static int nvgpu_tsg_bind_scheduling_domain(struct nvgpu_tsg *tsg,
+		struct nvgpu_tsg_bind_scheduling_domain_args *args)
+{
+
+	return nvgpu_tsg_bind_domain(tsg, args->domain_name);
+}
+
 #ifdef CONFIG_NVGPU_CHANNEL_TSG_CONTROL
 static int gk20a_tsg_get_event_data_from_id(struct nvgpu_tsg *tsg,
 				unsigned int event_id,
@@ -814,6 +821,20 @@ long nvgpu_ioctl_tsg_dev_ioctl(struct file *filp, unsigned int cmd,
 		break;
 		}
 
+	case NVGPU_TSG_IOCTL_BIND_SCHEDULING_DOMAIN:
+		{
+		err = gk20a_busy(g);
+		if (err) {
+			nvgpu_err(g,
+			   "failed to host gk20a for ioctl cmd: 0x%x", cmd);
+			break;
+		}
+		err = nvgpu_tsg_bind_scheduling_domain(tsg,
+				(struct nvgpu_tsg_bind_scheduling_domain_args *)buf);
+		gk20a_idle(g);
+		break;
+		}
+
 	case NVGPU_IOCTL_TSG_ENABLE:
 		{
 		err = gk20a_busy(g);

include/uapi/linux/nvgpu.h

@@ -38,6 +38,12 @@ struct nvgpu_tsg_bind_channel_ex_args {
 	__u8 reserved[16];
 };
 
+struct nvgpu_tsg_bind_scheduling_domain_args {
+	/* in: name of the domain this tsg will be bound to */
+	__u8 domain_name[16];
+	__u8 reserved[16];
+};
+
 /*
  * This struct helps to report the SM error state of a single SM.
  * This acts upon the currently resident TSG context.
@@ -131,11 +137,14 @@ struct nvgpu_tsg_set_l2_sector_promotion_args {
 #define NVGPU_TSG_IOCTL_SET_L2_SECTOR_PROMOTION \
 	_IOW(NVGPU_TSG_IOCTL_MAGIC, 15, \
 			struct nvgpu_tsg_set_l2_sector_promotion_args)
+#define NVGPU_TSG_IOCTL_BIND_SCHEDULING_DOMAIN \
+	_IOW(NVGPU_TSG_IOCTL_MAGIC, 16, \
+			struct nvgpu_tsg_bind_scheduling_domain_args)
 #define NVGPU_TSG_IOCTL_MAX_ARG_SIZE	\
-		sizeof(struct nvgpu_tsg_bind_channel_ex_args)
+		sizeof(struct nvgpu_tsg_bind_scheduling_domain_args)
 
 #define NVGPU_TSG_IOCTL_LAST		\
-	_IOC_NR(NVGPU_TSG_IOCTL_SET_L2_SECTOR_PROMOTION)
+	_IOC_NR(NVGPU_TSG_IOCTL_BIND_SCHEDULING_DOMAIN)
 
 /*
  * /dev/nvhost-dbg-gpu device