Add CZ support to queues module

Signed-off-by: Ben Goz <ben.goz@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>

1 files changed, 140 insertions, 6 deletions

diff --git a/src/queues.c b/src/queues.c
index 1feb9ba..2d7692f 100644
--- a/src/queues.c
+++ b/src/queues.c
@@ -30,15 +30,73 @@
 #include <sys/mman.h>
 #include <math.h>
 #include <stdio.h>
+#include <sys/types.h>
+#include <sys/mman.h>
+#include <fcntl.h>
 
 /* 1024 doorbells, 4 bytes each doorbell */
 #define DOORBELLS_PAGE_SIZE	1024 * 4
 
+struct device_info
+{
+	uint32_t ctx_save_restore_size;
+	uint32_t eop_buffer_size;
+};
+
+struct device_info kaveri_device_info = {
+	.ctx_save_restore_size = 0,
+	.eop_buffer_size = 0,
+};
+
+struct device_info carrizo_device_info = {
+	.ctx_save_restore_size = 2756608,
+	.eop_buffer_size = 4096,
+};
+
+struct device_id
+{
+	uint16_t dev_id;
+	struct device_info *dev_info;
+};
+
+struct device_id supported_devices[] = {
+	{ 0x1304, &kaveri_device_info },	/* Kaveri */
+	{ 0x1305, &kaveri_device_info },	/* Kaveri */
+	{ 0x1306, &kaveri_device_info },	/* Kaveri */
+	{ 0x1307, &kaveri_device_info },	/* Kaveri */
+	{ 0x1309, &kaveri_device_info },	/* Kaveri */
+	{ 0x130A, &kaveri_device_info },	/* Kaveri */
+	{ 0x130B, &kaveri_device_info },	/* Kaveri */
+	{ 0x130C, &kaveri_device_info },	/* Kaveri */
+	{ 0x130D, &kaveri_device_info },	/* Kaveri */
+	{ 0x130E, &kaveri_device_info },	/* Kaveri */
+	{ 0x130F, &kaveri_device_info },	/* Kaveri */
+	{ 0x1310, &kaveri_device_info },	/* Kaveri */
+	{ 0x1311, &kaveri_device_info },	/* Kaveri */
+	{ 0x1312, &kaveri_device_info },	/* Kaveri */
+	{ 0x1313, &kaveri_device_info },	/* Kaveri */
+	{ 0x1315, &kaveri_device_info },	/* Kaveri */
+	{ 0x1316, &kaveri_device_info },	/* Kaveri */
+	{ 0x1317, &kaveri_device_info },	/* Kaveri */
+	{ 0x1318, &kaveri_device_info },	/* Kaveri */
+	{ 0x131B, &kaveri_device_info },	/* Kaveri */
+	{ 0x131C, &kaveri_device_info },	/* Kaveri */
+	{ 0x131D, &kaveri_device_info },	/* Kaveri */
+	{ 0x9870, &carrizo_device_info },	/* Carrizo */
+	{ 0x9874, &carrizo_device_info },	/* Carrizo */
+	{ 0x9875, &carrizo_device_info },	/* Carrizo */
+	{ 0x9876, &carrizo_device_info },	/* Carrizo */
+	{ 0x9877, &carrizo_device_info },	/* Carrizo */
+	{ 0, NULL }
+};
+
 struct queue
 {
 	uint32_t queue_id;
 	uint32_t wptr;
 	uint32_t rptr;
+	void *eop_buffer;
+	void *ctx_save_restore;
 };
 
 struct process_doorbells
@@ -50,6 +108,74 @@ struct process_doorbells
 
 struct process_doorbells doorbells[] = {[0 ... (NUM_OF_SUPPORTED_GPUS-1)] = {.need_mmap = true, .doorbells = NULL, .doorbells_mutex = PTHREAD_MUTEX_INITIALIZER}};
 
+static struct device_info *get_device_info_by_dev_id(uint16_t dev_id)
+{
+	int i = 0;
+	while (supported_devices[i].dev_id != 0) {
+		if (supported_devices[i].dev_id == dev_id) {
+			return supported_devices[i].dev_info;
+		}
+		i++;
+	}
+
+	return NULL;
+}
+
+static void free_queue(struct queue *q)
+{
+	if (q->eop_buffer)
+		free(q->eop_buffer);
+	if (q->ctx_save_restore)
+		free(q->ctx_save_restore);
+	free(q);
+}
+
+static void* allocate_exec_aligned_memory(uint32_t size, uint32_t align)
+{
+	void *ptr;
+	int retval;
+
+	retval = posix_memalign(&ptr, align, size);
+	if (retval != 0)
+		return NULL;
+
+	retval = mprotect(ptr, size, PROT_READ | PROT_WRITE | PROT_EXEC);
+	if (retval != 0) {
+		free(ptr);
+		return NULL;
+	}
+
+	memset(ptr, 0, size);
+	return ptr;
+}
+
+static int handle_concrete_asic(struct device_info *dev_info, struct queue *q,
+								struct kfd_ioctl_create_queue_args *args)
+{
+	if (dev_info) {
+		if (dev_info->eop_buffer_size > 0) {
+			q->eop_buffer =
+					allocate_exec_aligned_memory(dev_info->eop_buffer_size, PAGE_SIZE);
+			if (q->eop_buffer == NULL) {
+				return HSAKMT_STATUS_NO_MEMORY;
+			}
+			args->eop_buffer_address = (uintptr_t)q->eop_buffer;
+			args->eop_buffer_size = dev_info->eop_buffer_size;
+		}
+		if (dev_info->ctx_save_restore_size > 0) {
+			args->ctx_save_restore_size = dev_info->ctx_save_restore_size;
+			q->ctx_save_restore =
+					allocate_exec_aligned_memory(dev_info->ctx_save_restore_size, PAGE_SIZE);
+			if (q->ctx_save_restore == NULL) {;
+				return HSAKMT_STATUS_NO_MEMORY;
+			}
+			args->ctx_save_restore_address = (uintptr_t)q->ctx_save_restore;
+		}
+	}
+
+	return HSAKMT_STATUS_SUCCESS;
+}
+
 HSAKMT_STATUS
 HSAKMTAPI
 hsaKmtCreateQueue(
@@ -65,6 +191,8 @@ hsaKmtCreateQueue(
 {
 	HSAKMT_STATUS result;
 	uint32_t gpu_id;
+	uint16_t dev_id;
+	struct device_info *dev_info;
 	int err;
 	void* ptr;
 	CHECK_KFD_OPEN();
@@ -75,23 +203,28 @@ hsaKmtCreateQueue(
 
 	struct queue *q = malloc(sizeof(struct queue));
 	if (q == NULL)
-	{
 		return HSAKMT_STATUS_NO_MEMORY;
-	}
-
 	memset(q, 0, sizeof(*q));
 
 	struct kfd_ioctl_create_queue_args args;
 	memset(&args, 0, sizeof(args));
 
+	dev_id = get_device_id_by_node(NodeId);
+	dev_info = get_device_info_by_dev_id(dev_id);
 	args.gpu_id = gpu_id;
 
+	err = handle_concrete_asic(dev_info, q, &args);
+	if (err != HSAKMT_STATUS_SUCCESS) {
+		free_queue(q);
+		return err;
+	}
+
 	switch (Type)
 	{
 	case HSA_QUEUE_COMPUTE: args.queue_type = KFD_IOC_QUEUE_TYPE_COMPUTE; break;
 	case HSA_QUEUE_SDMA: free(q); return HSAKMT_STATUS_NOT_IMPLEMENTED;
 	case HSA_QUEUE_COMPUTE_AQL: args.queue_type = KFD_IOC_QUEUE_TYPE_COMPUTE_AQL; break;
-	default: free(q); return HSAKMT_STATUS_INVALID_PARAMETER;
+	default: free_queue(q); return HSAKMT_STATUS_INVALID_PARAMETER;
 	}
 
 	if (Type != HSA_QUEUE_COMPUTE_AQL)
@@ -111,7 +244,7 @@ hsaKmtCreateQueue(
 
 	if (err == -1)
 	{
-		free(q);
+		free_queue(q);
 		return HSAKMT_STATUS_ERROR;
 	}
 
@@ -126,6 +259,7 @@ hsaKmtCreateQueue(
 		if (ptr == MAP_FAILED) {
 			pthread_mutex_unlock(&doorbells[NodeId].doorbells_mutex);
 			hsaKmtDestroyQueue(q->queue_id);
+			free_queue(q);
 			return HSAKMT_STATUS_ERROR;
 		}
 
@@ -201,7 +335,7 @@ hsaKmtDestroyQueue(
 	}
 	else
 	{
-		free(q);
+		free_queue(q);
 		return HSAKMT_STATUS_SUCCESS;
 	}
 }