gpu: nvgpu: Add subctx programming for MIG

This CL covers the following code changes,
1) Added api to init inst_block for more than one subctxs.
2) Added logic to limit the subctx bind based on
   max. VEID count allocated to a gr instance.
3) Renamed nvgpu_grmgr_get_gr_runlist_id.

JIRA NVGPU-5647

Change-Id: Ifec8164a9e5f46fbd0538c3dd50e19ee63667a54
Signed-off-by: Lakshmanan M <lm@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvgpu/+/2418463
Reviewed-by: automaticguardword <automaticguardword@nvidia.com>
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
Reviewed-by: svc-mobile-cert <svc-mobile-cert@nvidia.com>
Reviewed-by: Rajesh Devaraj <rdevaraj@nvidia.com>
Reviewed-by: Dinesh T <dt@nvidia.com>
Reviewed-by: Vaibhav Kachore <vkachore@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/channel.c

@@ -141,9 +141,20 @@ void nvgpu_channel_commit_va(struct nvgpu_channel *c)
 
 	nvgpu_log_fn(g, " ");
 
+	if (g->ops.mm.init_inst_block_for_subctxs != NULL) {
+		u32 subctx_count = nvgpu_channel_get_max_subctx_count(c);
+
+		nvgpu_log(g, gpu_dbg_info | gpu_dbg_mig,
+			"chid: %d max_subctx_count[%u] ",
+			c->chid, subctx_count);
+		g->ops.mm.init_inst_block_for_subctxs(&c->inst_block, c->vm,
+				c->vm->gmmu_page_sizes[GMMU_PAGE_SIZE_BIG],
+				subctx_count);
+	} else {
 		g->ops.mm.init_inst_block(&c->inst_block, c->vm,
 				c->vm->gmmu_page_sizes[GMMU_PAGE_SIZE_BIG]);
 	}
+}
 
 int nvgpu_channel_update_runlist(struct nvgpu_channel *c, bool add)
 {

drivers/gpu/nvgpu/common/grmgr/grmgr.c

@@ -390,7 +390,8 @@ bool nvgpu_grmgr_is_valid_runlist_id(struct gk20a *g,
 	return false;
 }
 
-u32 nvgpu_grmgr_get_gr_runlist_id(struct gk20a *g, u32 gpu_instance_id)
+u32 nvgpu_grmgr_get_gpu_instance_runlist_id(struct gk20a *g,
+		u32 gpu_instance_id)
 {
 	if (gpu_instance_id < g->mig.num_gpu_instances) {
 		struct nvgpu_gpu_instance *gpu_instance =

drivers/gpu/nvgpu/hal/fifo/ramfc_gv11b_fusa.c

@@ -53,7 +53,8 @@ int gv11b_ramfc_setup(struct nvgpu_channel *ch, u64 gpfifo_base,
 	nvgpu_log_info(g, "%llu %u", pbdma_acquire_timeout,
 		g->ops.pbdma.acquire_val(pbdma_acquire_timeout));
 
-	g->ops.ramin.init_subctx_pdb(g, mem, ch->vm->pdb.mem, replayable);
+	g->ops.ramin.init_subctx_pdb(g, mem, ch->vm->pdb.mem,
+		replayable, nvgpu_channel_get_max_subctx_count(ch));
 
 	nvgpu_mem_wr32(g, mem, ram_fc_gp_base_w(),
 		g->ops.pbdma.get_gp_base(gpfifo_base));

drivers/gpu/nvgpu/hal/fifo/ramin_gv11b.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2016-2020, NVIDIA CORPORATION.  All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
@@ -32,7 +32,7 @@ void gv11b_ramin_set_gr_ptr(struct gk20a *g,
 		struct nvgpu_mem *inst_block, u64 gpu_va);
 void gv11b_ramin_init_subctx_pdb(struct gk20a *g,
 		struct nvgpu_mem *inst_block, struct nvgpu_mem *pdb_mem,
-		bool replayable);
+		bool replayable, u32 max_subctx_count);
 void gv11b_ramin_set_eng_method_buffer(struct gk20a *g,
 		struct nvgpu_mem *inst_block, u64 gpu_va);
 

drivers/gpu/nvgpu/hal/fifo/ramin_gv11b_fusa.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015-2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2015-2020, NVIDIA CORPORATION.  All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
@@ -46,27 +46,45 @@ void gv11b_ramin_set_gr_ptr(struct gk20a *g,
 }
 
 static void gv11b_subctx_commit_valid_mask(struct gk20a *g,
-		struct nvgpu_mem *inst_block)
+		struct nvgpu_mem *inst_block, u32 max_subctx_count)
 {
 	u32 id;
+	u32 subctx_count = max_subctx_count;
+
+	for (id = 0U; id < max_subctx_count; id += 32U) {
+		u32 subctx_mask_max_bit = ((subctx_count < 32U) ?
+			(subctx_count % 32U) : 0U);
+		u32 subctx_mask = U32_MAX;
+
+		if (subctx_mask_max_bit != 0U) {
+			subctx_mask = nvgpu_safe_sub_u32(
+				BIT32(subctx_mask_max_bit), 1U);
+		}
 
-	/* Make all subctx pdbs valid */
-	for (id = 0U; id < ram_in_sc_pdb_valid__size_1_v(); id += 32U) {
 		nvgpu_mem_wr32(g, inst_block,
-				ram_in_sc_pdb_valid_long_w(id), U32_MAX);
+				ram_in_sc_pdb_valid_long_w(id), subctx_mask);
+
+		nvgpu_log(g, gpu_dbg_info | gpu_dbg_mig,
+			"id[%d] max_subctx_count[%u] subctx_mask_max_bit[%u] "
+				"subctx_count[%u] subctx_mask[%x] ",
+			id, max_subctx_count, subctx_mask_max_bit,
+			subctx_count, subctx_mask);
+
+		if (subctx_count > 32U) {
+			subctx_count = nvgpu_safe_sub_u32(subctx_count, 32U);
+		}
 	}
 }
 
 static void gv11b_subctx_commit_pdb(struct gk20a *g,
 		struct nvgpu_mem *inst_block, struct nvgpu_mem *pdb_mem,
-		bool replayable)
+		bool replayable, u32 max_subctx_count)
 {
 	u32 lo, hi;
 	u32 subctx_id = 0;
 	u32 format_word;
 	u32 pdb_addr_lo, pdb_addr_hi;
 	u64 pdb_addr;
-	u32 max_subctx_count = ram_in_sc_page_dir_base_target__size_1_v();
 	u32 aperture = nvgpu_aperture_mask(g, pdb_mem,
 				ram_in_sc_page_dir_base_target_sys_mem_ncoh_v(),
 				ram_in_sc_page_dir_base_target_sys_mem_coh_v(),
@@ -100,10 +118,11 @@ static void gv11b_subctx_commit_pdb(struct gk20a *g,
 
 void gv11b_ramin_init_subctx_pdb(struct gk20a *g,
 		struct nvgpu_mem *inst_block, struct nvgpu_mem *pdb_mem,
-		bool replayable)
+		bool replayable, u32 max_subctx_count)
 {
-	gv11b_subctx_commit_pdb(g, inst_block, pdb_mem, replayable);
-	gv11b_subctx_commit_valid_mask(g, inst_block);
+	gv11b_subctx_commit_pdb(g, inst_block, pdb_mem, replayable,
+		max_subctx_count);
+	gv11b_subctx_commit_valid_mask(g, inst_block, max_subctx_count);
 
 }
 

drivers/gpu/nvgpu/hal/init/hal_gv11b.c

@@ -1053,6 +1053,7 @@ static const struct gops_mm gv11b_ops_mm = {
 	.setup_hw = nvgpu_mm_setup_hw,
 	.is_bar1_supported = gv11b_mm_is_bar1_supported,
 	.init_inst_block = gv11b_mm_init_inst_block,
+	.init_inst_block_for_subctxs = gv11b_mm_init_inst_block_for_subctxs,
 	.init_bar2_vm = gp10b_mm_init_bar2_vm,
 	.remove_bar2_vm = gp10b_mm_remove_bar2_vm,
 	.bar1_map_userd = NULL,

drivers/gpu/nvgpu/hal/init/hal_tu104.c

@@ -1104,6 +1104,7 @@ static const struct gops_mm tu104_ops_mm = {
 	.setup_hw = nvgpu_mm_setup_hw,
 	.is_bar1_supported = gv11b_mm_is_bar1_supported,
 	.init_inst_block = gv11b_mm_init_inst_block,
+	.init_inst_block_for_subctxs = gv11b_mm_init_inst_block_for_subctxs,
 	.init_bar2_vm = gp10b_mm_init_bar2_vm,
 	.remove_bar2_vm = gp10b_mm_remove_bar2_vm,
 	.get_flush_retries = tu104_mm_get_flush_retries,

drivers/gpu/nvgpu/hal/mm/mm_gv11b.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2019-2020, NVIDIA CORPORATION.  All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
@@ -30,6 +30,8 @@ struct vm_gk20a;
 
 void gv11b_mm_init_inst_block(struct nvgpu_mem *inst_block, struct vm_gk20a *vm,
 			      u32 big_page_size);
+void gv11b_mm_init_inst_block_for_subctxs(struct nvgpu_mem *inst_block,
+		struct vm_gk20a *vm, u32 big_page_size, u32 max_subctx_count);
 bool gv11b_mm_is_bar1_supported(struct gk20a *g);
 
 #endif

drivers/gpu/nvgpu/hal/mm/mm_gv11b_fusa.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2019-2020, NVIDIA CORPORATION.  All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
@@ -42,7 +42,30 @@ void gv11b_mm_init_inst_block(struct nvgpu_mem *inst_block,
 	}
 
 	if (g->ops.ramin.init_subctx_pdb != NULL) {
-		g->ops.ramin.init_subctx_pdb(g, inst_block, vm->pdb.mem, false);
+		g->ops.ramin.init_subctx_pdb(g, inst_block, vm->pdb.mem, false,
+			1U);
+	}
+}
+
+void gv11b_mm_init_inst_block_for_subctxs(struct nvgpu_mem *inst_block,
+		struct vm_gk20a *vm, u32 big_page_size, u32 max_subctx_count)
+{
+	struct gk20a *g = gk20a_from_vm(vm);
+	u64 pdb_addr = nvgpu_pd_gpu_addr(g, &vm->pdb);
+
+	nvgpu_log_info(g, "inst block phys = 0x%llx, kv = 0x%p",
+		nvgpu_inst_block_addr(g, inst_block), inst_block->cpu_va);
+
+	g->ops.ramin.init_pdb(g, inst_block, pdb_addr, vm->pdb.mem);
+
+	if ((big_page_size != 0U) &&
+			(g->ops.ramin.set_big_page_size != NULL)) {
+		g->ops.ramin.set_big_page_size(g, inst_block, big_page_size);
+	}
+
+	if (g->ops.ramin.init_subctx_pdb != NULL) {
+		g->ops.ramin.init_subctx_pdb(g, inst_block, vm->pdb.mem, false,
+			max_subctx_count);
 	}
 }
 

drivers/gpu/nvgpu/hal/vgpu/init/vgpu_hal_gv11b.c

@@ -756,6 +756,7 @@ static const struct gops_mm vgpu_gv11b_ops_mm = {
 	.setup_hw = NULL,
 	.is_bar1_supported = gv11b_mm_is_bar1_supported,
 	.init_inst_block = gv11b_mm_init_inst_block,
+	.init_inst_block_for_subctxs = gv11b_mm_init_inst_block_for_subctxs,
 	.init_bar2_vm = gp10b_mm_init_bar2_vm,
 	.remove_bar2_vm = gp10b_mm_remove_bar2_vm,
 	.bar1_map_userd = vgpu_mm_bar1_map_userd,

drivers/gpu/nvgpu/include/nvgpu/channel.h

@@ -1178,4 +1178,11 @@ static inline void nvgpu_channel_set_wdt_debug_dump(struct nvgpu_channel *ch,
 		bool dump) {}
 #endif
 
+/**
+ * @brief Get maximum sub context count.
+ *
+ * @param ch [in]	Channel pointer.
+ */
+u32 nvgpu_channel_get_max_subctx_count(struct nvgpu_channel *ch);
+
 #endif

drivers/gpu/nvgpu/include/nvgpu/gops/mm.h

@@ -540,6 +540,23 @@ struct gops_mm {
 	void (*init_inst_block)(struct nvgpu_mem *inst_block,
 			struct vm_gk20a *vm, u32 big_page_size);
 
+	/**
+	 * @brief HAL to initialize the instance block memory.
+	 * (for more than one subctx)
+	 *
+	 * @param inst_block    [in]	Pointer to instance block memory.
+	 * @param vm            [in]	Pointer to virtual memory context.
+	 * @param big_page_size [in]	Big page size supported by GMMU.
+	 * @param max_subctx_count [in] Max number of sub context.
+	 *
+	 * Initializes the instance block memory:
+	 * - Configures the pdb base, big page size and
+	 *   sub context's pdb base in context's instance block memory.
+	 */
+	void (*init_inst_block_for_subctxs)(struct nvgpu_mem *inst_block,
+			struct vm_gk20a *vm, u32 big_page_size,
+			u32 max_subctx_count);
+
 	/**
 	 * @brief HAL to get the maximum flush retry counts.
 	 *

drivers/gpu/nvgpu/include/nvgpu/gops/ramin.h

@@ -98,6 +98,7 @@ struct gops_ramin {
 	 * @param pdb_mem [in]		Memory descriptor of PDB.
 	 * @param replayable [in]	Indicates if errors are replayable
 	 * 				for this Instance Block.
+	 * @param max_subctx_count [in] Max number of sub context.
 	 *
 	 * This HAL configures PDB for all sub-contexts of Instance Block:
 	 * - Get max number of sub-contexts from HW.
@@ -116,7 +117,7 @@ struct gops_ramin {
 	void (*init_subctx_pdb)(struct gk20a *g,
 			struct nvgpu_mem *inst_block,
 			struct nvgpu_mem *pdb_mem,
-			bool replayable);
+			bool replayable, u32 max_subctx_count);
 
 	/**
 	 * @brief Instance Block shift.

drivers/gpu/nvgpu/include/nvgpu/grmgr.h

@@ -42,7 +42,8 @@ u32 nvgpu_grmgr_get_gr_gpc_phys_id(struct gk20a *g, u32 gr_instance_id,
 u32 nvgpu_grmgr_get_gr_instance_id(struct gk20a *g, u32 gpu_instance_id);
 bool nvgpu_grmgr_is_valid_runlist_id(struct gk20a *g,
 		u32 gpu_instance_id, u32 runlist_id);
-u32 nvgpu_grmgr_get_gr_runlist_id(struct gk20a *g, u32 gpu_instance_id);
+u32 nvgpu_grmgr_get_gpu_instance_runlist_id(struct gk20a *g,
+		u32 gpu_instance_id);
 u32 nvgpu_grmgr_get_gr_instance_id_for_syspipe(struct gk20a *g,
 		u32 gr_syspipe_id);
 u32 nvgpu_grmgr_get_gpu_instance_max_veid_count(struct gk20a *g,

drivers/gpu/nvgpu/os/linux/linux-channel.c

@@ -22,6 +22,7 @@
 #include <nvgpu/channel.h>
 #include <nvgpu/dma.h>
 #include <nvgpu/fence.h>
+#include <nvgpu/grmgr.h>
 
 /*
  * This is required for nvgpu_vm_find_buf() which is used in the tracing
@@ -628,6 +629,14 @@ u32 nvgpu_get_gpfifo_entry_size(void)
 	return sizeof(struct nvgpu_gpfifo_entry);
 }
 
+u32 nvgpu_channel_get_max_subctx_count(struct nvgpu_channel *ch)
+{
+	struct gk20a *g = ch->g;
+
+	return nvgpu_grmgr_get_gpu_instance_max_veid_count(g,
+		0U);
+}
+
 #ifdef CONFIG_DEBUG_FS
 static void trace_write_pushbuffer(struct nvgpu_channel *c,
 				   struct nvgpu_gpfifo_entry *g)

drivers/gpu/nvgpu/os/posix/posix-channel.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2018-2020, NVIDIA CORPORATION.  All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
@@ -30,3 +30,8 @@ u32 nvgpu_get_gpfifo_entry_size(void)
 	 */
 	return 8;
 }
+
+u32 nvgpu_channel_get_max_subctx_count(struct nvgpu_channel *ch)
+{
+	return 64;
+}

libs/dgpu/libnvgpu-drv-dgpu_safe.export

@@ -402,6 +402,7 @@ nvgpu_free_fixed
 nvgpu_free_gr_ctx_struct
 nvgpu_get
 nvgpu_get_gpfifo_entry_size
+nvgpu_channel_get_max_subctx_count
 nvgpu_get_pte
 nvgpu_gmmu_default_big_page_size
 nvgpu_gmmu_init_page_table

libs/igpu/libnvgpu-drv-igpu_safe.export

@@ -416,6 +416,7 @@ nvgpu_free_fixed
 nvgpu_free_gr_ctx_struct
 nvgpu_get
 nvgpu_get_gpfifo_entry_size
+nvgpu_channel_get_max_subctx_count
 nvgpu_get_pte
 nvgpu_gmmu_default_big_page_size
 nvgpu_gmmu_init_page_table

userspace/required_tests.ini

@@ -501,7 +501,7 @@ test_channel_suspend_resume_serviceable_chs.suspend_resume=0
 test_channel_sw_quiesce.sw_quiesce=0
 test_fifo_init_support.init_support=0
 test_fifo_remove_support.remove_support=0
-test_nvgpu_channel_commit_va.channel_commit_va=0
+test_nvgpu_channel_commit_va.channel_commit_va=2
 test_nvgpu_get_gpfifo_entry_size.get_gpfifo_entry_size=0
 test_trace_write_pushbuffers.trace_write_pushbuffers=0
 
@@ -1041,7 +1041,7 @@ test_gm20b_ramin_set_big_page_size.set_big_page_size=0
 test_gp10b_ramin_init_pdb.init_pdb=0
 
 [ramin_gv11b_fusa]
-test_gv11b_ramin_init_subctx_pdb.init_subctx_pdb=0
+test_gv11b_ramin_init_subctx_pdb.init_subctx_pdb=2
 test_gv11b_ramin_set_eng_method_buffer.set_eng_method_buf=0
 test_gv11b_ramin_set_gr_ptr.set_gr_ptr=0
 
@@ -1083,7 +1083,7 @@ test_map_buffer_error_cases.map_buffer_error_cases=0
 test_nvgpu_vm_alloc_va.nvgpu_vm_alloc_va=0
 test_vm_area_error_cases.vm_area_error_cases=0
 test_vm_aspace_id.vm_aspace_id=0
-test_vm_bind.vm_bind=0
+test_vm_bind.vm_bind=2
 test_gk20a_from_vm.gk20a_from_vm=0
 test_vm_pde_coverage_bit_count.vm_pde_coverage_bit_count=0
 test_nvgpu_insert_mapped_buf.nvgpu_insert_mapped_buf=0

userspace/units/fifo/channel/nvgpu-channel.c

@@ -1979,7 +1979,7 @@ struct unit_module_test nvgpu_channel_tests[] = {
 	UNIT_TEST(channel_put_warn, test_channel_put_warn, &unit_ctx, 0),
 	UNIT_TEST(referenceable_cleanup, test_ch_referenceable_cleanup, &unit_ctx, 0),
 	UNIT_TEST(abort_cleanup, test_channel_abort_cleanup, &unit_ctx, 0),
-	UNIT_TEST(channel_commit_va, test_nvgpu_channel_commit_va, &unit_ctx, 0),
+	UNIT_TEST(channel_commit_va, test_nvgpu_channel_commit_va, &unit_ctx, 2),
 	UNIT_TEST(get_gpfifo_entry_size, test_nvgpu_get_gpfifo_entry_size, &unit_ctx, 0),
 	UNIT_TEST(trace_write_pushbuffers, test_trace_write_pushbuffers, &unit_ctx, 0),
 	UNIT_TEST(remove_support, test_fifo_remove_support, &unit_ctx, 0),

userspace/units/fifo/ramfc/gv11b/nvgpu-ramfc-gv11b.c

@@ -133,7 +133,7 @@ static int stub_ramfc_commit_userd(struct nvgpu_channel *ch)
 
 static void stub_ramin_init_subctx_pdb(struct gk20a *g,
 			struct nvgpu_mem *inst_block, struct nvgpu_mem *pdb_mem,
-			bool replayable)
+			bool replayable, u32 max_subctx_count)
 {
 	global_count++;
 }

userspace/units/fifo/ramin/gv11b/ramin-gv11b-fusa.c

@@ -147,7 +147,7 @@ int test_gv11b_ramin_init_subctx_pdb(struct unit_module *m, struct gk20a *g,
 		}
 
 		gv11b_ramin_init_subctx_pdb(g, &inst_block, &pdb_mem,
-								replayable);
+								replayable, 64);
 
 		for (subctx_id = 0; subctx_id < max_subctx_count; subctx_id++) {
 			addr_lo = ram_in_sc_page_dir_base_vol_w(subctx_id);
@@ -215,7 +215,7 @@ done:
 
 struct unit_module_test ramin_gv11b_fusa_tests[] = {
 	UNIT_TEST(set_gr_ptr, test_gv11b_ramin_set_gr_ptr, NULL, 0),
-	UNIT_TEST(init_subctx_pdb, test_gv11b_ramin_init_subctx_pdb, NULL, 0),
+	UNIT_TEST(init_subctx_pdb, test_gv11b_ramin_init_subctx_pdb, NULL, 2),
 	UNIT_TEST(set_eng_method_buf, test_gv11b_ramin_set_eng_method_buffer, NULL, 0),
 };
 

userspace/units/mm/vm/vm.c

@@ -2076,7 +2076,7 @@ struct unit_module_test vm_tests[] = {
 	UNIT_TEST(init_error_paths, test_init_error_paths, NULL, 0),
 	UNIT_TEST(map_buffer_error_cases, test_map_buffer_error_cases, NULL, 0),
 	UNIT_TEST(nvgpu_vm_alloc_va, test_nvgpu_vm_alloc_va, NULL, 0),
-	UNIT_TEST(vm_bind, test_vm_bind, NULL, 0),
+	UNIT_TEST(vm_bind, test_vm_bind, NULL, 2),
 	UNIT_TEST(vm_aspace_id, test_vm_aspace_id, NULL, 0),
 	UNIT_TEST(vm_area_error_cases, test_vm_area_error_cases, NULL, 0),
 	UNIT_TEST_REQ("NVGPU-RQCD-45.C2",