gpu: nvgpu: gsp: sched: domain management apis

Changes:
- Added Domain management APIs with interfaces to communicate with
  GSP scheduler.
- Domain creation shall be done inside NVGPU and respective Domain
  and runlist info are sent to GSP for scheduling.

Design: https://confluence.nvidia.com/display/TGS/GSP+Scheduler+Interface+Specifications

NVGPU-7371

Signed-off-by: Ramesh Mylavarapu <rmylavarapu@nvidia.com>
Change-Id: Icba7f1ed3b9b2f409aac346084dd9a123c9d3779
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvgpu/+/2682686
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
This commit is contained in:
Ramesh Mylavarapu
2022-04-06 17:56:20 +05:30
committed by mobile promotions
parent 5828594de1
commit 951ad46819
7 changed files with 384 additions and 105 deletions

View File

@@ -20,8 +20,6 @@
* DEALINGS IN THE SOFTWARE.
*/
#include "gsp_runlist.h"
#include <nvgpu/gk20a.h>
#include <nvgpu/log.h>
#include <nvgpu/gsp.h>
@@ -33,25 +31,60 @@
#include "ipc/gsp_cmd.h"
#include "ipc/gsp_msg.h"
#include "gsp_runlist.h"
static void gsp_handle_cmd_ack(struct gk20a *g, struct nv_flcn_msg_gsp *msg,
void *param, u32 status)
{
bool *command_ack = param;
nvgpu_gsp_dbg(g, " ");
nvgpu_log_fn(g, " ");
switch (msg->hdr.unit_id) {
case NV_GSP_UNIT_NULL:
nvgpu_info(g, "Reply to NV_GSP_UNIT_NULL");
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_NULL");
*command_ack = true;
break;
case NV_GSP_UNIT_SUBMIT_RUNLIST:
nvgpu_info(g, "Reply to NV_GSP_UNIT_RUNLIST_INFO");
case NV_GSP_UNIT_DOMAIN_SUBMIT:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_DOMAIN_SUBMIT");
*command_ack = true;
break;
case NV_GSP_UNIT_DEVICES_INFO:
nvgpu_info(g, "Reply to NV_GSP_UNIT_DEVICES_INFO");
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_DEVICES_INFO");
*command_ack = true;
break;
case NV_GSP_UNIT_DOMAIN_ADD:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_DOMAIN_ADD");
*command_ack = true;
break;
case NV_GSP_UNIT_DOMAIN_DELETE:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_DOMAIN_DELETE");
*command_ack = true;
break;
case NV_GSP_UNIT_DOMAIN_UPDATE:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_DOMAIN_UPDATE");
*command_ack = true;
break;
case NV_GSP_UNIT_RUNLIST_UPDATE:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_RUNLIST_UPDATE");
*command_ack = true;
break;
case NV_GSP_UNIT_START_SCHEDULER:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_START_SCHEDULER");
*command_ack = true;
break;
case NV_GSP_UNIT_STOP_SCHEDULER:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_STOP_SCHEDULER");
*command_ack = true;
break;
case NV_GSP_UNIT_QUERY_NO_OF_DOMAINS:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_QUERY_NO_OF_DOMAINS");
g->gsp_sched->no_of_domains = msg->msg.no_of_domains.no_of_domains;
*command_ack = true;
break;
case NV_GSP_UNIT_QUERY_ACTIVE_DOMAIN:
nvgpu_gsp_dbg(g, "Reply to NV_GSP_UNIT_QUERY_ACTIVE_DOMAIN");
g->gsp_sched->active_domain = msg->msg.active_domain.active_domain;
*command_ack = true;
break;
default:
@@ -63,73 +96,59 @@ static void gsp_handle_cmd_ack(struct gk20a *g, struct nv_flcn_msg_gsp *msg,
(void)status;
}
static void gsp_get_runlist_info(struct gk20a *g,
struct nvgpu_gsp_runlist_info *rl_info, struct nvgpu_runlist *runlist)
static int gsp_send_cmd_and_wait_for_ack(struct gk20a *g,
struct nv_flcn_cmd_gsp *cmd, u32 unit_id, u32 size)
{
u64 runlist_iova;
u32 aperture, num_entries;
runlist_iova = nvgpu_mem_get_addr(g, &runlist->domain->mem_hw->mem);
num_entries = runlist->domain->mem_hw->count;
aperture = g->ops.runlist.get_runlist_aperture(g, runlist);
rl_info->runlist_base_lo = u64_lo32(runlist_iova);
rl_info->runlist_base_hi = u64_hi32(runlist_iova);
rl_info->aperture = aperture;
rl_info->num_entries = num_entries;
rl_info->runlist_id = runlist->id;
}
int nvgpu_gsp_runlist_submit(struct gk20a *g, struct nvgpu_runlist *runlist)
{
struct nv_flcn_cmd_gsp cmd;
bool command_ack = false;
int err = 0;
size_t tmp_size;
nvgpu_gsp_dbg(g, " ");
(void) memset(&cmd, 0, sizeof(struct nv_flcn_cmd_gsp));
cmd.hdr.unit_id = NV_GSP_UNIT_SUBMIT_RUNLIST;
tmp_size = GSP_CMD_HDR_SIZE + sizeof(struct nvgpu_gsp_runlist_info);
tmp_size = GSP_CMD_HDR_SIZE + size;
nvgpu_assert(tmp_size <= U64(U8_MAX));
cmd.hdr.size = (u8)tmp_size;
cmd->hdr.size = tmp_size;
cmd->hdr.unit_id = unit_id;
/* copy domain info into cmd buffer */
gsp_get_runlist_info(g, &cmd.cmd.runlist, runlist);
err = nvgpu_gsp_cmd_post(g, &cmd, GSP_NV_CMDQ_LOG_ID,
err = nvgpu_gsp_cmd_post(g, cmd, GSP_NV_CMDQ_LOG_ID,
gsp_handle_cmd_ack, &command_ack, U32_MAX);
if (err != 0) {
nvgpu_err(g, "command post failed");
nvgpu_err(g, "cmd post failed unit_id:0x%x", unit_id);
goto exit;
}
err = nvgpu_gsp_wait_message_cond(g, nvgpu_get_poll_timeout(g),
&command_ack, U8(true));
if (err != 0) {
nvgpu_err(g, "command ack receive failed");
nvgpu_err(g, "cmd ack receive failed unit_id:0x%x", unit_id);
}
exit:
return err;
}
static void gsp_get_device_info(struct gk20a *g,
struct nvgpu_gsp_device_info *dev_info)
static int gsp_get_async_ce(struct gk20a *g, struct nvgpu_device *device,
u32 instance)
{
const struct nvgpu_device *device;
const struct nvgpu_device *lces[NVGPU_MIG_MAX_ENGINES] = { };
u32 num_lce;
/* Only GRAPHICS 0-instance is supported by GSP scheduler.
* In future, more devices can be looped through and send it to the GSP.
*/
device = nvgpu_device_get(g, NVGPU_DEVTYPE_GRAPHICS, 0);
nvgpu_assert(device != NULL);
num_lce = nvgpu_device_get_async_copies(g, lces, NVGPU_MIG_MAX_ENGINES);
if (num_lce == 0) {
nvgpu_err(g, "Async CEs not supported");
return -1;
}
*device = *lces[instance];
return 0;
}
static void gsp_get_device_info(struct gk20a *g, u8 device_id,
struct nvgpu_gsp_device_info *dev_info,
const struct nvgpu_device *device)
{
/* copy domain info into cmd buffer */
dev_info->device_id = NVGPU_DEVTYPE_GRAPHICS;
dev_info->device_id = device_id;
dev_info->is_engine = true;
dev_info->engine_type = device->type;
dev_info->engine_id = device->engine_id;
@@ -139,37 +158,208 @@ static void gsp_get_device_info(struct gk20a *g,
dev_info->runlist_pri_base = device->rl_pri_base;
}
int nvgpu_gsp_send_devices_info(struct gk20a *g)
static int gsp_sched_send_devices_info(struct gk20a *g,
u8 device_id, const struct nvgpu_device *device)
{
struct nv_flcn_cmd_gsp cmd;
bool command_ack = false;
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
size_t tmp_size;
nvgpu_gsp_dbg(g, " ");
(void) memset(&cmd, 0, sizeof(struct nv_flcn_cmd_gsp));
cmd.hdr.unit_id = NV_GSP_UNIT_DEVICES_INFO;
tmp_size = GSP_CMD_HDR_SIZE + sizeof(struct nvgpu_gsp_device_info);
nvgpu_assert(tmp_size <= U64(U8_MAX));
cmd.hdr.size = (u8)tmp_size;
/* copy domain info into cmd buffer */
gsp_get_device_info(g, &cmd.cmd.device);
gsp_get_device_info(g, device_id, &cmd.cmd.device, device);
err = nvgpu_gsp_cmd_post(g, &cmd, GSP_NV_CMDQ_LOG_ID,
gsp_handle_cmd_ack, &command_ack, U32_MAX);
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_DEVICES_INFO, sizeof(struct nvgpu_gsp_device_info));
return err;
}
int nvgpu_gsp_sched_send_devices_info(struct gk20a *g)
{
const struct nvgpu_device *gr_dev = NULL;
struct nvgpu_device ce_dev = { };
u8 instance = 0;
int err = 0;
/*
* Only GR0 is supported
*/
gr_dev = nvgpu_device_get(g, NVGPU_DEVTYPE_GRAPHICS, instance);
if (gr_dev == NULL) {
nvgpu_err(g, "Get GR0 device info failed");
goto exit;
}
err = gsp_sched_send_devices_info(g,
GSP_SCHED_GR0_DEVICE_ID, gr_dev);
if (err != 0) {
nvgpu_err(g, "command post failed");
nvgpu_err(g, "send GR0 device info failed");
goto exit;
}
err = nvgpu_gsp_wait_message_cond(g, nvgpu_get_poll_timeout(g),
&command_ack, U8(true));
/*
* Only Async CE0 is supported
*/
err = gsp_get_async_ce(g, &ce_dev, instance);
if (err != 0) {
nvgpu_err(g, "command ack receive failed");
nvgpu_err(g, "Get Async CE0 device info failed");
goto exit;
}
err = gsp_sched_send_devices_info(g,
GSP_SCHED_ASYNC_CE0_DEVICE_ID, &ce_dev);
if (err != 0) {
nvgpu_err(g, "send Async CE0 device info failed");
goto exit;
}
exit:
return err;
}
int nvgpu_gsp_sched_domain_add(struct gk20a *g,
struct nvgpu_gsp_domain_info *gsp_dom)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
/* copy domain info into cmd buffer */
cmd.cmd.domain = *gsp_dom;
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_DOMAIN_ADD, sizeof(struct nvgpu_gsp_domain_info));
return err;
}
int nvgpu_gsp_sched_domain_update(struct gk20a *g,
struct nvgpu_gsp_domain_info *gsp_dom)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
/* copy domain info into cmd buffer */
cmd.cmd.domain = *gsp_dom;
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_DOMAIN_UPDATE, sizeof(struct nvgpu_gsp_domain_info));
return err;
}
int nvgpu_gsp_sched_domain_delete(struct gk20a *g, u32 domain_id)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
/* copy domain_id to cmd buffer */
cmd.cmd.domain_id.domain_id = domain_id;
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_DOMAIN_DELETE, sizeof(struct nvgpu_gsp_domain_id));
return err;
}
int nvgpu_gsp_sched_domain_submit(struct gk20a *g, u32 domain_id)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
/* copy domain_id to cmd buffer */
cmd.cmd.domain_id.domain_id = domain_id;
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_DOMAIN_SUBMIT, sizeof(struct nvgpu_gsp_domain_id));
return err;
}
int nvgpu_gsp_sched_runlist_update(struct gk20a *g,
struct nvgpu_gsp_runlist_info *gsp_rl)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
/* copy domain info into cmd buffer */
cmd.cmd.runlist = *gsp_rl;
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_RUNLIST_UPDATE, sizeof(struct nvgpu_gsp_runlist_info));
return err;
}
int nvgpu_gsp_sched_query_no_of_domains(struct gk20a *g, u32 *no_of_domains)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_QUERY_NO_OF_DOMAINS, 0);
if (err != 0) {
nvgpu_err(g, "send cmd failed");
return err;
}
*no_of_domains = g->gsp_sched->no_of_domains;
return err;
}
int nvgpu_gsp_sched_query_active_domain(struct gk20a *g, u32 *active_domain)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_QUERY_ACTIVE_DOMAIN, 0);
if (err != 0) {
nvgpu_err(g, "send cmd failed");
return err;
}
*active_domain = g->gsp_sched->active_domain;
return err;
}
int nvgpu_gsp_sched_start(struct gk20a *g)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_START_SCHEDULER, 0);
if (err != 0) {
nvgpu_err(g, "send cmd failed");
}
return err;
}
int nvgpu_gsp_sched_stop(struct gk20a *g)
{
struct nv_flcn_cmd_gsp cmd = { };
int err = 0;
nvgpu_gsp_dbg(g, " ");
err = gsp_send_cmd_and_wait_for_ack(g, &cmd,
NV_GSP_UNIT_STOP_SCHEDULER, 0);
return err;
}

View File

@@ -23,9 +23,9 @@
#ifndef NVGPU_GSP_RUNLIST
#define NVGPU_GSP_RUNLIST
#include <nvgpu/device.h>
#define GSP_SCHED_GR0_DEVICE_ID 0U
#define GSP_SCHED_ASYNC_CE0_DEVICE_ID 1U
#define NVGPU_GSP_MAX_DEVTYPE 1U
struct nvgpu_gsp_device_info {
/*
@@ -65,36 +65,10 @@ struct nvgpu_gsp_device_info {
u32 dev_pri_base;
};
struct nvgpu_gsp_runlist_info {
struct nvgpu_gsp_domain_id {
/*
* Device id to which this runlist belongs to
* Domain id
*/
u8 device_id;
/*
* Domain id to which this runlist need to mapped to
*/
u8 domain_id;
/*
* Indicates how many runlist entries are in the newly submitted runlist
*/
u32 num_entries;
/*
* Indicates how many runlist aperture
*/
u32 aperture;
/*
* ID contains the identifier of the runlist.
*/
u32 runlist_id;
/*
*NV_RUNLIST_SUBMIT_BASE_L0 in-memory location of runlist.
*/
u32 runlist_base_lo;
/*
*NV_RUNLIST_SUBMIT_BASE_Hi in-memory location of runlist.
*/
u32 runlist_base_hi;
u32 domain_id;
};
int nvgpu_gsp_send_devices_info(struct gk20a *g);
#endif // NVGPU_GSP_RUNLIST

View File

@@ -204,9 +204,10 @@ int nvgpu_gsp_sched_bootstrap_ns(struct gk20a *g)
goto de_init;
}
status = nvgpu_gsp_send_devices_info(g);
status = nvgpu_gsp_sched_send_devices_info(g);
if (status != 0) {
nvgpu_err(g, "gsp send device info failed");
goto de_init;
}
return status;

View File

@@ -44,6 +44,10 @@ struct nvgpu_gsp_sched {
/* set to true once init received */
bool gsp_ready;
/* Log info of Domains */
u32 no_of_domains;
u32 active_domain;
};
#endif /* GSP_SCHEDULER_H */

View File

@@ -24,21 +24,30 @@
#define NVGPU_GSP_CMD_IF_H
#include <nvgpu/types.h>
#include <nvgpu/gsp_sched.h>
#include "../gsp_runlist.h"
#include "gsp_seq.h"
struct gk20a;
#define GSP_NV_CMDQ_LOG_ID 0U
#define GSP_NV_CMDQ_LOG_ID__LAST 0U
#define GSP_NV_MSGQ_LOG_ID 1U
#define GSP_NV_CMDQ_LOG_ID 0U
#define GSP_NV_CMDQ_LOG_ID__LAST 0U
#define GSP_NV_MSGQ_LOG_ID 1U
#define NV_GSP_UNIT_REWIND NV_FLCN_UNIT_ID_REWIND
#define NV_GSP_UNIT_NULL 0x01U
#define NV_GSP_UNIT_INIT 0x02U
#define NV_GSP_UNIT_DEVICES_INFO 0x03U
#define NV_GSP_UNIT_SUBMIT_RUNLIST 0x04U
#define NV_GSP_UNIT_END 0x0AU
#define NV_GSP_UNIT_REWIND NV_FLCN_UNIT_ID_REWIND
#define NV_GSP_UNIT_NULL 0x01U
#define NV_GSP_UNIT_INIT 0x02U
#define NV_GSP_UNIT_DEVICES_INFO 0x03U
#define NV_GSP_UNIT_DOMAIN_SUBMIT 0x04U
#define NV_GSP_UNIT_DOMAIN_ADD 0x05U
#define NV_GSP_UNIT_DOMAIN_DELETE 0x06U
#define NV_GSP_UNIT_DOMAIN_UPDATE 0x07U
#define NV_GSP_UNIT_RUNLIST_UPDATE 0x08U
#define NV_GSP_UNIT_START_SCHEDULER 0x09U
#define NV_GSP_UNIT_STOP_SCHEDULER 0x0AU
#define NV_GSP_UNIT_QUERY_NO_OF_DOMAINS 0x0BU
#define NV_GSP_UNIT_QUERY_ACTIVE_DOMAIN 0X0CU
#define NV_GSP_UNIT_END 0x0DU
#define GSP_MSG_HDR_SIZE U32(sizeof(struct gsp_hdr))
#define GSP_CMD_HDR_SIZE U32(sizeof(struct gsp_hdr))
@@ -53,8 +62,10 @@ struct gsp_hdr {
struct nv_flcn_cmd_gsp {
struct gsp_hdr hdr;
union {
struct nvgpu_gsp_domain_id domain_id;
struct nvgpu_gsp_device_info device;
struct nvgpu_gsp_runlist_info runlist;
struct nvgpu_gsp_domain_info domain;
} cmd;
};

View File

@@ -61,11 +61,23 @@ union nv_flcn_msg_gsp_init {
struct gsp_init_msg_gsp_init gsp_init;
};
struct nvgpu_gsp_domain_msg_no_of_domains {
u8 msg_type;
u32 no_of_domains;
};
struct nvgpu_gsp_domain_msg_active_domain {
u8 msg_type;
u32 active_domain;
};
struct nv_flcn_msg_gsp {
struct gsp_hdr hdr;
union {
union nv_flcn_msg_gsp_init init;
struct nvgpu_gsp_domain_msg_no_of_domains no_of_domains;
struct nvgpu_gsp_domain_msg_active_domain active_domain;
} msg;
};

View File

@@ -25,9 +25,96 @@
struct gk20a;
struct nvgpu_gsp_sched;
/*
* Scheduler shall support only two engines with two runlists per domain.
* 1. GR0
* 2. Async CE0
*/
#define TOTAL_NO_OF_RUNLISTS 2U
struct nvgpu_gsp_runlist_info {
/*
* Is this runlist valid, this field will be updated by NVGPU which tell GSP
* to submit this runlist or ignore for that domain.
*/
bool is_runlist_valid;
/*
* This is don't care for KMD.
*/
bool is_runlist_updated;
/*
* Device id to which this runlist belongs to.
*/
u8 device_id;
/*
* Domain Id to which this runlist belongs to.
*/
u32 domain_id;
/*
* ID contains the identifier of the runlist.
*/
u32 runlist_id;
/*
* Indicates how many runlist entries are in the runlist.
*/
u32 num_entries;
/*
* Indicates runlist memory aperture.
*/
u32 aperture;
/*
*NV_RUNLIST_SUBMIT_BASE_L0 in-memory location of runlist.
*/
u32 runlist_base_lo;
/*
*NV_RUNLIST_SUBMIT_BASE_Hi in-memory location of runlist.
*/
u32 runlist_base_hi;
};
struct nvgpu_gsp_domain_info {
/*
* Is the current Domain Active. This is don't care for KMD.
*/
bool is_domain_active;
/*
* Is the current Domain Valid. This is don't care for KMD.
*/
bool is_domain_valid;
/*
* Domain Id
*/
u32 domain_id;
/*
* Priority of the Domain for priority driven scheduling.
*/
u32 priority;
/*
* Time-slicing of the domain for which scheduler will schedule it for.
*/
u32 time_slicing;
/*
* Runlist info
*/
struct nvgpu_gsp_runlist_info runlist_info[TOTAL_NO_OF_RUNLISTS];
};
int nvgpu_gsp_sched_bootstrap_ns(struct gk20a *g);
int nvgpu_gsp_sched_sw_init(struct gk20a *g);
void nvgpu_gsp_sched_sw_deinit(struct gk20a *g);
void nvgpu_gsp_sched_suspend(struct gk20a *g, struct nvgpu_gsp_sched *gsp_sched);
void nvgpu_gsp_sched_isr(struct gk20a *g);
int nvgpu_gsp_sched_send_devices_info(struct gk20a *g);
int nvgpu_gsp_sched_domain_submit(struct gk20a *g, u32 domain_id);
int nvgpu_gsp_sched_domain_add(struct gk20a *g,
struct nvgpu_gsp_domain_info *gsp_dom);
int nvgpu_gsp_sched_domain_update(struct gk20a *g,
struct nvgpu_gsp_domain_info *gsp_dom);
int nvgpu_gsp_sched_runlist_update(struct gk20a *g,
struct nvgpu_gsp_runlist_info *gsp_rl);
int nvgpu_gsp_sched_domain_delete(struct gk20a *g, u32 domain_id);
int nvgpu_gsp_sched_query_active_domain(struct gk20a *g, u32 *active_domain);
int nvgpu_gsp_sched_query_no_of_domains(struct gk20a *g, u32 *no_of_domains);
int nvgpu_gsp_sched_start(struct gk20a *g);
int nvgpu_gsp_sched_stop(struct gk20a *g);
#endif /* GSP_SCHED_H */