[PATCH 09/13] drm/amdkfd: Add support for VI in DQM

[Oded Gabbay]

From: Ben Goz <ben.goz at amd.com>

This patch adds support for the VI APU in the DQM module.

Most of the functionality of DQM is shared between CI and VI. Therefore,
only a handful of functions are required to be in the
H/W-specific part of DQM.

Signed-off-by: Ben Goz <ben.goz at amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay at gmail.com>
---
 .../drm/amd/amdkfd/kfd_device_queue_manager_vi.c   | 103 ++++++++++++++++++++-
 1 file changed, 99 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_vi.c
index 4c15212..44c38e8 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_vi.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_vi.c
@@ -22,6 +22,10 @@
  */
 
 #include "kfd_device_queue_manager.h"
+#include "gca/gfx_8_0_enum.h"
+#include "gca/gfx_8_0_sh_mask.h"
+#include "gca/gfx_8_0_enum.h"
+#include "oss/oss_3_0_sh_mask.h"
 
 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
 				   struct qcm_process_device *qpd,
@@ -37,14 +41,40 @@ static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
 
 void device_queue_manager_init_vi(struct device_queue_manager_asic_ops *ops)
 {
-	pr_warn("amdkfd: VI DQM is not currently supported\n");
-
 	ops->set_cache_memory_policy = set_cache_memory_policy_vi;
 	ops->register_process = register_process_vi;
 	ops->initialize = initialize_cpsch_vi;
 	ops->init_sdma_vm = init_sdma_vm;
 }
 
+static uint32_t compute_sh_mem_bases_64bit(unsigned int top_address_nybble)
+{
+	/* In 64-bit mode, we can only control the top 3 bits of the LDS,
+	 * scratch and GPUVM apertures.
+	 * The hardware fills in the remaining 59 bits according to the
+	 * following pattern:
+	 * LDS:		X0000000'00000000 - X0000001'00000000 (4GB)
+	 * Scratch:	X0000001'00000000 - X0000002'00000000 (4GB)
+	 * GPUVM:	Y0010000'00000000 - Y0020000'00000000 (1TB)
+	 *
+	 * (where X/Y is the configurable nybble with the low-bit 0)
+	 *
+	 * LDS and scratch will have the same top nybble programmed in the
+	 * top 3 bits of SH_MEM_BASES.PRIVATE_BASE.
+	 * GPUVM can have a different top nybble programmed in the
+	 * top 3 bits of SH_MEM_BASES.SHARED_BASE.
+	 * We don't bother to support different top nybbles
+	 * for LDS/Scratch and GPUVM.
+	 */
+
+	BUG_ON((top_address_nybble & 1) || top_address_nybble > 0xE ||
+		top_address_nybble == 0);
+
+	return top_address_nybble << 12 |
+			(top_address_nybble << 12) <<
+			SH_MEM_BASES__SHARED_BASE__SHIFT;
+}
+
 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
 				   struct qcm_process_device *qpd,
 				   enum cache_policy default_policy,
@@ -52,18 +82,83 @@ static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
 				   void __user *alternate_aperture_base,
 				   uint64_t alternate_aperture_size)
 {
-	return false;
+	uint32_t default_mtype;
+	uint32_t ape1_mtype;
+
+	default_mtype = (default_policy == cache_policy_coherent) ?
+			MTYPE_CC :
+			MTYPE_NC;
+
+	ape1_mtype = (alternate_policy == cache_policy_coherent) ?
+			MTYPE_CC :
+			MTYPE_NC;
+
+	qpd->sh_mem_config = (qpd->sh_mem_config &
+			SH_MEM_CONFIG__ADDRESS_MODE_MASK) |
+		SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
+				SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
+		default_mtype << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
+		ape1_mtype << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
+		SH_MEM_CONFIG__PRIVATE_ATC_MASK;
+
+	return true;
 }
 
 static int register_process_vi(struct device_queue_manager *dqm,
 					struct qcm_process_device *qpd)
 {
-	return -1;
+	struct kfd_process_device *pdd;
+	unsigned int temp;
+
+	BUG_ON(!dqm || !qpd);
+
+	pdd = qpd_to_pdd(qpd);
+
+	/* check if sh_mem_config register already configured */
+	if (qpd->sh_mem_config == 0) {
+		qpd->sh_mem_config =
+			SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
+				SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
+			MTYPE_CC << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
+			MTYPE_CC << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
+			SH_MEM_CONFIG__PRIVATE_ATC_MASK;
+
+		qpd->sh_mem_ape1_limit = 0;
+		qpd->sh_mem_ape1_base = 0;
+	}
+
+	if (qpd->pqm->process->is_32bit_user_mode) {
+		temp = get_sh_mem_bases_32(pdd);
+		qpd->sh_mem_bases = temp << SH_MEM_BASES__SHARED_BASE__SHIFT;
+		qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA32 <<
+					SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
+	} else {
+		temp = get_sh_mem_bases_nybble_64(pdd);
+		qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
+		qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA64 <<
+			SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
+	}
+
+	pr_debug("kfd: is32bit process: %d sh_mem_bases nybble: 0x%X and register 0x%X\n",
+		qpd->pqm->process->is_32bit_user_mode, temp, qpd->sh_mem_bases);
+
+	return 0;
 }
 
 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
 				struct qcm_process_device *qpd)
 {
+	uint32_t value = (1 << SDMA0_RLC0_VIRTUAL_ADDR__ATC__SHIFT);
+
+	if (q->process->is_32bit_user_mode)
+		value |= (1 << SDMA0_RLC0_VIRTUAL_ADDR__PTR32__SHIFT) |
+				get_sh_mem_bases_32(qpd_to_pdd(qpd));
+	else
+		value |= ((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
+				SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &&
+				SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
+
+	q->properties.sdma_vm_addr = value;
 }
 
 static int initialize_cpsch_vi(struct device_queue_manager *dqm)
-- 
2.4.3