Files
android_kernel_samsung_sm86…/msm/eva/cvp_power.c
George Shen 304e2cf9c0 msm: eva: Allocate 1MB debug buffer for FW
For FW to hold debug info. Simplify driver structure for easy
dump in T32.

Change-Id: Ib310a3d9fe3437d5ce49783eb813fbb2d8bd3216
Signed-off-by: George Shen <quic_sqiao@quicinc.com>
2023-06-02 14:40:35 -07:00

550 行
15 KiB
C

/* SPDX-License-Identifier: GPL-2.0-only
*
* Copyright (c) 2022-2023, Qualcomm Innovation Center, Inc. All rights reserved.
* Copyright (c) 2018-2021, The Linux Foundation. All rights reserved.
*/
#include "msm_cvp.h"
#include "cvp_power.h"
static inline int find_max(unsigned long *array, unsigned int num)
{
int i, max = 0;
for (i = 0; i < num; i++)
max = array[i] > max ? array[i] : max;
return max;
}
static bool is_subblock_profile_existed(struct msm_cvp_inst *inst)
{
return (inst->prop.cycles[HFI_HW_OD] ||
inst->prop.cycles[HFI_HW_MPU] ||
inst->prop.cycles[HFI_HW_FDU] ||
inst->prop.cycles[HFI_HW_ICA] ||
inst->prop.cycles[HFI_HW_VADL] ||
inst->prop.cycles[HFI_HW_TOF] ||
inst->prop.cycles[HFI_HW_RGE] ||
inst->prop.cycles[HFI_HW_XRA] ||
inst->prop.cycles[HFI_HW_LSR]);
}
static char hw_names[HFI_MAX_HW_THREADS][8] = {{"FDU"}, {"OD"}, {"MPU"}, {"ICA"},
{"VADL"}, {"TOF"}, {"RGE"}, {"XRA"},
{"LSR"}};
static void aggregate_power_update(struct msm_cvp_core *core,
struct cvp_power_level *nrt_pwr,
struct cvp_power_level *rt_pwr,
unsigned int max_clk_rate)
{
struct msm_cvp_inst *inst;
int i, j;
unsigned long blocks_sum[2][HFI_MAX_HW_THREADS] = {0};
unsigned long fw_sum[2] = {0}, max_cycle[2] = {0}, op_max_cycle[2] = {0};
unsigned long op_blocks_max[2][HFI_MAX_HW_THREADS] = {0};
unsigned long op_fw_max[2] = {0}, bw_sum[2] = {0}, op_bw_max[2] = {0};
for (j = 0; j < HFI_MAX_HW_THREADS; j++)
core->dyn_clk.sum_fps[j] = 0;
list_for_each_entry(inst, &core->instances, list) {
if (inst->state == MSM_CVP_CORE_INVALID ||
inst->state == MSM_CVP_CORE_UNINIT ||
!is_subblock_profile_existed(inst))
continue;
if (inst->prop.priority <= CVP_RT_PRIO_THRESHOLD) {
/* Non-realtime session use index 0 */
i = 0;
} else {
i = 1;
}
for (j = 0; j < HFI_MAX_HW_THREADS; j++)
if (inst->prop.cycles[j])
dprintk(CVP_PWR, "pwrUpdate %s %u\n",
hw_names[j], inst->prop.cycles[j]);
for (j = 0; j < HFI_MAX_HW_THREADS; j++)
if (inst->prop.op_cycles[j])
dprintk(CVP_PWR, "pwrUpdate_OP %s %u\n",
hw_names[j], inst->prop.op_cycles[j]);
dprintk(CVP_PWR, " fw %u fw_o %u\n", inst->prop.fw_cycles,
inst->prop.fw_op_cycles);
for (j = 0; j < HFI_MAX_HW_THREADS; j++)
blocks_sum[i][j] += inst->prop.cycles[j];
fw_sum[i] += inst->prop.fw_cycles;
for (j = 0; j < HFI_MAX_HW_THREADS; j++)
op_blocks_max[i][j] =
(op_blocks_max[i][j] >= inst->prop.op_cycles[j]) ?
op_blocks_max[i][j] : inst->prop.op_cycles[j];
op_fw_max[i] =
(op_fw_max[i] >= inst->prop.fw_op_cycles) ?
op_fw_max[i] : inst->prop.fw_op_cycles;
bw_sum[i] += inst->prop.ddr_bw;
op_bw_max[i] =
(op_bw_max[i] >= inst->prop.ddr_op_bw) ?
op_bw_max[i] : inst->prop.ddr_op_bw;
for (j = 0; j < HFI_MAX_HW_THREADS; j++) {
if (inst->prop.fps[j])
dprintk(CVP_PWR, "fps %s %d ", hw_names[j],
inst->prop.fps[j]);
core->dyn_clk.sum_fps[j] += inst->prop.fps[j];
}
for (j = 0; j < HFI_MAX_HW_THREADS; j++)
if (core->dyn_clk.sum_fps[j])
dprintk(CVP_PWR, "sum_fps %s %d ", hw_names[j],
core->dyn_clk.sum_fps[j]);
}
for (i = 0; i < 2; i++) {
max_cycle[i] = find_max(&blocks_sum[i][0], HFI_MAX_HW_THREADS);
op_max_cycle[i] = find_max(&op_blocks_max[i][0], HFI_MAX_HW_THREADS);
op_max_cycle[i] =
(op_max_cycle[i] > max_clk_rate) ?
max_clk_rate : op_max_cycle[i];
bw_sum[i] = (bw_sum[i] >= op_bw_max[i]) ?
bw_sum[i] : op_bw_max[i];
}
nrt_pwr->core_sum += max_cycle[0];
nrt_pwr->op_core_sum = (nrt_pwr->op_core_sum >= op_max_cycle[0]) ?
nrt_pwr->op_core_sum : op_max_cycle[0];
nrt_pwr->bw_sum += bw_sum[0];
rt_pwr->core_sum += max_cycle[1];
rt_pwr->op_core_sum = (rt_pwr->op_core_sum >= op_max_cycle[1]) ?
rt_pwr->op_core_sum : op_max_cycle[1];
rt_pwr->bw_sum += bw_sum[1];
}
/**
* adjust_bw_freqs(): calculate CVP clock freq and bw required to sustain
* required use case.
* Bandwidth vote will be best-effort, not returning error if the request
* b/w exceeds max limit.
* Clock vote from non-realtime sessions will be best effort, not returning
* error if the aggreated session clock request exceeds max limit.
* Clock vote from realtime session will be hard request. If aggregated
* session clock request exceeds max limit, the function will return
* error.
*
* Ensure caller acquires clk_lock!
*/
static int adjust_bw_freqs(void)
{
struct msm_cvp_core *core;
struct iris_hfi_device *hdev;
struct bus_info *bus = NULL;
struct clock_set *clocks;
struct clock_info *cl;
struct allowed_clock_rates_table *tbl = NULL;
unsigned int tbl_size;
unsigned int cvp_min_rate, cvp_max_rate, max_bw = 0, min_bw = 0;
struct cvp_power_level rt_pwr = {0}, nrt_pwr = {0};
unsigned long tmp, core_sum, op_core_sum, bw_sum;
int i, rc = 0, bus_count = 0;
unsigned long ctrl_freq;
core = cvp_driver->cvp_core;
hdev = core->device->hfi_device_data;
clocks = &core->resources.clock_set;
cl = &clocks->clock_tbl[clocks->count - 1];
tbl = core->resources.allowed_clks_tbl;
tbl_size = core->resources.allowed_clks_tbl_size;
cvp_min_rate = tbl[0].clock_rate;
cvp_max_rate = tbl[tbl_size - 1].clock_rate;
for (bus_count = 0; bus_count < core->resources.bus_set.count; bus_count++) {
if (!strcmp(core->resources.bus_set.bus_tbl[bus_count].name, "cvp-ddr")) {
bus = &core->resources.bus_set.bus_tbl[bus_count];
max_bw = bus->range[1];
min_bw = max_bw/10;
}
}
if (!bus) {
dprintk(CVP_ERR, "bus node is NULL for cvp-ddr\n");
return -EINVAL;
}
aggregate_power_update(core, &nrt_pwr, &rt_pwr, cvp_max_rate);
dprintk(CVP_PWR, "PwrUpdate nrt %u %u rt %u %u\n",
nrt_pwr.core_sum, nrt_pwr.op_core_sum,
rt_pwr.core_sum, rt_pwr.op_core_sum);
if (rt_pwr.core_sum > cvp_max_rate) {
dprintk(CVP_WARN, "%s clk vote out of range %lld\n",
__func__, rt_pwr.core_sum);
return -ENOTSUPP;
}
core_sum = rt_pwr.core_sum + nrt_pwr.core_sum;
op_core_sum = (rt_pwr.op_core_sum >= nrt_pwr.op_core_sum) ?
rt_pwr.op_core_sum : nrt_pwr.op_core_sum;
core_sum = (core_sum >= op_core_sum) ?
core_sum : op_core_sum;
if (core_sum > cvp_max_rate) {
core_sum = cvp_max_rate;
} else if (core_sum <= cvp_min_rate) {
core_sum = cvp_min_rate;
} else {
for (i = 1; i < tbl_size; i++)
if (core_sum <= tbl[i].clock_rate)
break;
core_sum = tbl[i].clock_rate;
}
bw_sum = rt_pwr.bw_sum + nrt_pwr.bw_sum;
bw_sum = bw_sum >> 10;
bw_sum = (bw_sum > max_bw) ? max_bw : bw_sum;
bw_sum = (bw_sum < min_bw) ? min_bw : bw_sum;
dprintk(CVP_PWR, "%s %lld %lld\n", __func__,
core_sum, bw_sum);
if (!cl->has_scaling) {
dprintk(CVP_ERR, "Cannot scale CVP clock\n");
return -EINVAL;
}
tmp = core->curr_freq;
core->curr_freq = core_sum;
core->orig_core_sum = core_sum;
rc = msm_cvp_set_clocks(core);
if (rc) {
dprintk(CVP_ERR,
"Failed to set clock rate %u %s: %d %s\n",
core_sum, cl->name, rc, __func__);
core->curr_freq = tmp;
return rc;
}
ctrl_freq = (core->curr_freq*3)>>1;
core->dyn_clk.conf_freq = core->curr_freq;
for (i = 0; i < HFI_MAX_HW_THREADS; ++i) {
core->dyn_clk.hi_ctrl_lim[i] = core->dyn_clk.sum_fps[i] ?
ctrl_freq/core->dyn_clk.sum_fps[i] : 0;
core->dyn_clk.lo_ctrl_lim[i] =
core->dyn_clk.hi_ctrl_lim[i];
}
hdev->clk_freq = core->curr_freq;
rc = msm_cvp_set_bw(bus, bw_sum);
return rc;
}
int msm_cvp_update_power(struct msm_cvp_inst *inst)
{
int rc = 0;
struct msm_cvp_core *core;
struct msm_cvp_inst *s;
if (!inst) {
dprintk(CVP_ERR, "%s: invalid params\n", __func__);
return -EINVAL;
}
s = cvp_get_inst_validate(inst->core, inst);
if (!s)
return -ECONNRESET;
core = inst->core;
mutex_lock(&core->clk_lock);
rc = adjust_bw_freqs();
mutex_unlock(&core->clk_lock);
cvp_put_inst(s);
return rc;
}
static int cvp_readjust_clock(struct msm_cvp_core *core,
u32 avg_cycles, enum hfi_hw_thread i)
{
int rc = 0;
struct allowed_clock_rates_table *tbl = NULL;
unsigned int tbl_size = 0;
unsigned int cvp_min_rate = 0, cvp_max_rate = 0;
unsigned long tmp = core->curr_freq;
unsigned long lo_freq = 0;
u32 j;
tbl = core->resources.allowed_clks_tbl;
tbl_size = core->resources.allowed_clks_tbl_size;
cvp_min_rate = tbl[0].clock_rate;
cvp_max_rate = tbl[tbl_size - 1].clock_rate;
if (!((avg_cycles > core->dyn_clk.hi_ctrl_lim[i] &&
core->curr_freq != cvp_max_rate) ||
(avg_cycles <= core->dyn_clk.lo_ctrl_lim[i] &&
core->curr_freq != cvp_min_rate))) {
return rc;
}
core->curr_freq = ((avg_cycles * core->dyn_clk.sum_fps[i]) << 1)/3;
dprintk(CVP_PWR,
"%s - cycles tot %u, avg %u. sum_fps %u, cur_freq %u\n",
__func__,
core->dyn_clk.cycle[i].total,
avg_cycles,
core->dyn_clk.sum_fps[i],
core->curr_freq);
if (core->curr_freq > cvp_max_rate) {
core->curr_freq = cvp_max_rate;
lo_freq = (tbl_size > 1) ?
tbl[tbl_size - 2].clock_rate :
cvp_min_rate;
} else if (core->curr_freq <= cvp_min_rate) {
core->curr_freq = cvp_min_rate;
lo_freq = cvp_min_rate;
} else {
for (j = 1; j < tbl_size; j++)
if (core->curr_freq <= tbl[j].clock_rate)
break;
core->curr_freq = tbl[j].clock_rate;
lo_freq = tbl[j-1].clock_rate;
}
if (core->orig_core_sum > core->curr_freq) {
dprintk(CVP_PWR,
"%s - %d - Cancel readjust, core %u, freq %u\n",
__func__, i, core->orig_core_sum, core->curr_freq);
core->curr_freq = tmp;
return rc;
}
dprintk(CVP_PWR,
"%s:%d - %d - Readjust to %u\n",
__func__, __LINE__, i, core->curr_freq);
rc = msm_cvp_set_clocks(core);
if (rc) {
dprintk(CVP_ERR,
"Failed to set clock rate %u: %d %s\n",
core->curr_freq, rc, __func__);
core->curr_freq = tmp;
} else {
lo_freq = (lo_freq < core->dyn_clk.conf_freq) ?
core->dyn_clk.conf_freq : lo_freq;
core->dyn_clk.hi_ctrl_lim[i] = core->dyn_clk.sum_fps[i] ?
((core->curr_freq*3)>>1)/core->dyn_clk.sum_fps[i] : 0;
core->dyn_clk.lo_ctrl_lim[i] =
core->dyn_clk.sum_fps[i] ?
((lo_freq*3)>>1)/core->dyn_clk.sum_fps[i] : 0;
dprintk(CVP_PWR,
"%s - Readjust clk to %u. New lim [%d] hi %u lo %u\n",
__func__, core->curr_freq, i,
core->dyn_clk.hi_ctrl_lim[i],
core->dyn_clk.lo_ctrl_lim[i]);
}
return rc;
}
int cvp_check_clock(struct msm_cvp_inst *inst,
struct cvp_hfi_msg_session_hdr_ext *hdr)
{
int rc = 0;
u32 i, j;
u32 hw_cycles[HFI_MAX_HW_THREADS] = {0};
u32 fw_cycles = 0;
struct msm_cvp_core *core = inst->core;
for (i = 0; i < HFI_MAX_HW_ACTIVATIONS_PER_FRAME; ++i)
fw_cycles += hdr->fw_cycles[i];
for (i = 0; i < HFI_MAX_HW_THREADS; ++i)
for (j = 0; j < HFI_MAX_HW_ACTIVATIONS_PER_FRAME; ++j)
hw_cycles[i] += hdr->hw_cycles[i][j];
dprintk(CVP_PWR, "%s - cycles fw %u. FDU %d MPU %d ODU %d ICA %d\n",
__func__, fw_cycles, hw_cycles[0],
hw_cycles[1], hw_cycles[2], hw_cycles[3]);
mutex_lock(&core->clk_lock);
for (i = 0; i < HFI_MAX_HW_THREADS; ++i) {
dprintk(CVP_PWR, "%s - %d: hw_cycles %u, tens_thresh %u\n",
__func__, i, hw_cycles[i],
core->dyn_clk.hi_ctrl_lim[i]);
if (core->dyn_clk.hi_ctrl_lim[i]) {
if (core->dyn_clk.cycle[i].size < CVP_CYCLE_STAT_SIZE)
core->dyn_clk.cycle[i].size++;
else
core->dyn_clk.cycle[i].total -=
core->dyn_clk.cycle[i].busy[
core->dyn_clk.cycle[i].idx];
if (hw_cycles[i]) {
core->dyn_clk.cycle[i].busy[
core->dyn_clk.cycle[i].idx]
= hw_cycles[i] + fw_cycles;
core->dyn_clk.cycle[i].total
+= hw_cycles[i] + fw_cycles;
dprintk(CVP_PWR,
"%s: busy (hw + fw) cycles = %u\n",
__func__,
core->dyn_clk.cycle[i].busy[
core->dyn_clk.cycle[i].idx]);
dprintk(CVP_PWR, "total cycles %u\n",
core->dyn_clk.cycle[i].total);
} else {
core->dyn_clk.cycle[i].busy[
core->dyn_clk.cycle[i].idx] =
hdr->busy_cycles;
core->dyn_clk.cycle[i].total +=
hdr->busy_cycles;
dprintk(CVP_PWR,
"%s - busy cycles = %u total %u\n",
__func__,
core->dyn_clk.cycle[i].busy[
core->dyn_clk.cycle[i].idx],
core->dyn_clk.cycle[i].total);
}
core->dyn_clk.cycle[i].idx =
(core->dyn_clk.cycle[i].idx ==
CVP_CYCLE_STAT_SIZE-1) ?
0 : core->dyn_clk.cycle[i].idx+1;
dprintk(CVP_PWR, "%s - %d: size %u, tens_thresh %u\n",
__func__, i, core->dyn_clk.cycle[i].size,
core->dyn_clk.hi_ctrl_lim[i]);
if (core->dyn_clk.cycle[i].size == CVP_CYCLE_STAT_SIZE
&& core->dyn_clk.hi_ctrl_lim[i] != 0) {
u32 avg_cycles =
core->dyn_clk.cycle[i].total>>3;
rc = cvp_readjust_clock(core,
avg_cycles,
i);
}
}
}
mutex_unlock(&core->clk_lock);
return rc;
}
unsigned int msm_cvp_get_hw_aggregate_cycles(enum hfi_hw_thread hwblk)
{
struct msm_cvp_core *core;
struct msm_cvp_inst *inst;
unsigned long cycles_sum = 0;
core = cvp_driver->cvp_core;
if (!core) {
dprintk(CVP_ERR, "%s: invalid core\n", __func__);
return -EINVAL;
}
mutex_lock(&core->clk_lock);
list_for_each_entry(inst, &core->instances, list) {
if (inst->state == MSM_CVP_CORE_INVALID ||
inst->state == MSM_CVP_CORE_UNINIT ||
!is_subblock_profile_existed(inst))
continue;
switch (hwblk) {
case HFI_HW_FDU:
{
cycles_sum += inst->prop.cycles[HFI_HW_FDU];
break;
}
case HFI_HW_ICA:
{
cycles_sum += inst->prop.cycles[HFI_HW_ICA];
break;
}
case HFI_HW_MPU:
{
cycles_sum += inst->prop.cycles[HFI_HW_MPU];
break;
}
case HFI_HW_OD:
{
cycles_sum += inst->prop.cycles[HFI_HW_OD];
break;
}
case HFI_HW_VADL:
{
cycles_sum += inst->prop.cycles[HFI_HW_VADL];
break;
}
case HFI_HW_TOF:
{
cycles_sum += inst->prop.cycles[HFI_HW_TOF];
break;
}
case HFI_HW_RGE:
{
cycles_sum += inst->prop.cycles[HFI_HW_RGE];
break;
}
case HFI_HW_XRA:
{
cycles_sum += inst->prop.cycles[HFI_HW_XRA];
break;
}
case HFI_HW_LSR:
{
cycles_sum += inst->prop.cycles[HFI_HW_LSR];
break;
}
default:
dprintk(CVP_ERR, "unrecognized hw block %d\n",
hwblk);
break;
}
}
mutex_unlock(&core->clk_lock);
cycles_sum = cycles_sum&0xFFFFFFFF;
return (unsigned int)cycles_sum;
}
bool check_clock_required(struct msm_cvp_inst *inst,
struct eva_kmd_hfi_packet *hdr)
{
struct cvp_hfi_msg_session_hdr_ext *ehdr =
(struct cvp_hfi_msg_session_hdr_ext *)hdr;
bool clock_check = false;
if (!msm_cvp_dcvs_disable &&
ehdr->packet_type == HFI_MSG_SESSION_CVP_FD) {
if (ehdr->size == sizeof(struct cvp_hfi_msg_session_hdr_ext)
+ sizeof(struct cvp_hfi_buf_type)) {
struct msm_cvp_core *core = inst->core;
dprintk(CVP_PWR, "busy cycle %d, total %d\n",
ehdr->busy_cycles, ehdr->total_cycles);
if (core->dyn_clk.sum_fps[HFI_HW_FDU] ||
core->dyn_clk.sum_fps[HFI_HW_MPU] ||
core->dyn_clk.sum_fps[HFI_HW_OD] ||
core->dyn_clk.sum_fps[HFI_HW_ICA]) {
clock_check = true;
}
} else {
dprintk(CVP_WARN, "dcvs is disabled, %d != %d + %d\n",
ehdr->size, sizeof(struct cvp_hfi_msg_session_hdr_ext),
sizeof(struct cvp_hfi_buf_type));
}
}
return clock_check;
}