gpu: nvgpu: Simplify ref-counting on VMs

Simplify ref-counting on VMs: take a ref when a VM is bound to a
channel and drop a ref when a channel is freed.

Previously ref-counts were scattered over the driver. Also the CE
and CDE code would bind channels with custom rolled code. This was
because the gk20a_vm_bind_channel() function took an as_share as
the VM argument (the VM was then inferred from that as_share).
However, it is trivial to abtract that bit out and allow a central
bind channel function that just takes a VM and a channel.

Bug 1846718

Change-Id: I156aab259f6c7a2fa338408c6c4a3a464cd44a0c
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: http://git-master/r/1261886
Reviewed-by: Richard Zhao <rizhao@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>

drivers/gpu/nvgpu/gk20a/cde_gk20a.c

@@ -1203,9 +1203,7 @@ static int gk20a_cde_load(struct gk20a_cde_ctx *cde_ctx)
 	}
 
 	/* bind the channel to the vm */
-	gk20a_vm_get(&g->mm.cde.vm);
-	ch->vm = &g->mm.cde.vm;
-	err = channel_gk20a_commit_va(ch);
+	err = __gk20a_vm_bind_channel(&g->mm.cde.vm, ch);
 	if (err) {
 		gk20a_warn(cde_ctx->dev, "cde: could not bind vm");
 		goto err_commit_va;

drivers/gpu/nvgpu/gk20a/ce2_gk20a.c

@@ -453,9 +453,7 @@ u32 gk20a_ce_create_context_with_cb(struct device *dev,
 	 }
 
 	/* bind the channel to the vm */
-	gk20a_vm_get(&g->mm.ce.vm);
-	ce_ctx->vm = ce_ctx->ch->vm = &g->mm.ce.vm;
-	err = channel_gk20a_commit_va(ce_ctx->ch);
+	err = __gk20a_vm_bind_channel(&g->mm.ce.vm, ce_ctx->ch);
 	if (err) {
 		gk20a_err(ce_ctx->dev, "ce: could not bind vm");
 		goto end;

drivers/gpu/nvgpu/gk20a/channel_gk20a.c

@@ -1008,11 +1008,10 @@ static void gk20a_free_channel(struct channel_gk20a *ch, bool force)
 	if (ch->hw_sema)
 		gk20a_semaphore_free_hw_sema(ch);
 
-	/* release channel binding to the as_share */
-	if (ch_vm->as_share)
-		gk20a_as_release_share(ch_vm->as_share);
-	else
-		gk20a_vm_put(ch_vm);
+	/*
+	 * When releasing the channel we unbind the VM - so release the ref.
+	 */
+	gk20a_vm_put(ch_vm);
 
 	spin_lock(&ch->update_fn_lock);
 	ch->update_fn = NULL;
@@ -2252,14 +2251,11 @@ static int gk20a_channel_add_job(struct channel_gk20a *c,
 	int err = 0, num_mapped_buffers = 0;
 	bool pre_alloc_enabled = channel_gk20a_is_prealloc_enabled(c);
 
-	/* job needs reference to this vm (released in channel_update) */
-	gk20a_vm_get(vm);
-
 	if (!skip_buffer_refcounting) {
 		err = gk20a_vm_get_buffers(vm, &mapped_buffers,
 					&num_mapped_buffers);
 		if (err)
-			goto err_put_vm;
+			return err;
 	}
 
 	/* put() is done in gk20a_channel_update() when the job is done */
@@ -2293,8 +2289,6 @@ static int gk20a_channel_add_job(struct channel_gk20a *c,
 
 err_put_buffers:
 	gk20a_vm_put_buffers(vm, mapped_buffers, num_mapped_buffers);
-err_put_vm:
-	gk20a_vm_put(vm);
 
 	return err;
 }
@@ -2385,8 +2379,6 @@ static void gk20a_channel_clean_up_jobs(struct channel_gk20a *c,
 		gk20a_free_priv_cmdbuf(c, job->wait_cmd);
 		gk20a_free_priv_cmdbuf(c, job->incr_cmd);
 
-		/* job is done. release its vm reference (taken in add_job) */
-		gk20a_vm_put(vm);
 		/* another bookkeeping taken in add_job. caller must hold a ref
 		 * so this wouldn't get freed here. */
 		gk20a_channel_put(c);

drivers/gpu/nvgpu/gk20a/mm_gk20a.c

@@ -4199,10 +4199,8 @@ static int gk20a_init_sema_pool(struct vm_gk20a *vm)
 		return -ENOMEM;
 
 	vm->sema_pool = gk20a_semaphore_pool_alloc(sema_sea);
-	if (!vm->sema_pool) {
-		gk20a_vm_put(vm);
+	if (!vm->sema_pool)
 		return -ENOMEM;
-	}
 
 	/*
 	 * Allocate a chunk of GPU VA space for mapping the semaphores. We will
@@ -4227,7 +4225,7 @@ static int gk20a_init_sema_pool(struct vm_gk20a *vm)
 		gk20a_semaphore_pool_unmap(vm->sema_pool, vm);
 		nvgpu_free(vm->vma[gmmu_page_size_small],
 			   vm->sema_pool->gpu_va);
-		gk20a_vm_put(vm);
+		return err;
 	}
 
 	return 0;
@@ -4632,12 +4630,10 @@ int gk20a_vm_release_share(struct gk20a_as_share *as_share)
 	gk20a_dbg_fn("");
 
 	vm->as_share = NULL;
-
-	/* put as reference to vm */
-	gk20a_vm_put(vm);
-
 	as_share->vm = NULL;
 
+	gk20a_vm_put(vm);
+
 	return 0;
 }
 
@@ -4792,14 +4788,13 @@ int gk20a_vm_free_space(struct gk20a_as_share *as_share,
 	return err;
 }
 
-int gk20a_vm_bind_channel(struct gk20a_as_share *as_share,
-			  struct channel_gk20a *ch)
+int __gk20a_vm_bind_channel(struct vm_gk20a *vm, struct channel_gk20a *ch)
 {
 	int err = 0;
-	struct vm_gk20a *vm = as_share->vm;
 
 	gk20a_dbg_fn("");
 
+	gk20a_vm_get(vm);
 	ch->vm = vm;
 	err = channel_gk20a_commit_va(ch);
 	if (err)
@@ -4808,6 +4803,12 @@ int gk20a_vm_bind_channel(struct gk20a_as_share *as_share,
 	return err;
 }
 
+int gk20a_vm_bind_channel(struct gk20a_as_share *as_share,
+			  struct channel_gk20a *ch)
+{
+	return __gk20a_vm_bind_channel(as_share->vm, ch);
+}
+
 int gk20a_dmabuf_alloc_drvdata(struct dma_buf *dmabuf, struct device *dev)
 {
 	struct gk20a_dmabuf_priv *priv;

drivers/gpu/nvgpu/gk20a/mm_gk20a.h

@@ -707,6 +707,7 @@ int gk20a_vm_free_space(struct gk20a_as_share *as_share,
 			struct nvgpu_as_free_space_args *args);
 int gk20a_vm_bind_channel(struct gk20a_as_share *as_share,
 			  struct channel_gk20a *ch);
+int __gk20a_vm_bind_channel(struct vm_gk20a *vm, struct channel_gk20a *ch);
 
 /* batching eliminates redundant cache flushes and invalidates */
 void gk20a_vm_mapping_batch_start(struct vm_gk20a_mapping_batch *batch);

drivers/gpu/nvgpu/vgpu/mm_vgpu.c

@@ -446,6 +446,9 @@ static int vgpu_vm_bind_channel(struct gk20a_as_share *as_share,
 		err = -ENOMEM;
 	}
 
+	if (ch->vm)
+		gk20a_vm_get(ch->vm);
+
 	return err;
 }