SVS driver got accepted upstream but still has room to be improved.
Therefore, we add these patches to fix issues and coding style.
Change since v3:
- Remove cleanup codes in patch [01/14] for letting it backport easily
- Enclose svs debug cmd codes with "CONFIG_DEBUG_FS" where needed
- Reorder patch sequenece to put fix up patch first
Tested on:
MT8183 Platform (kernel v5.10)
MT8192 Platform (kernel v5.4)
Matthias Brugger (4):
soc: mediatek: mtk-svs: clean up platform probing
soc: mediatek: mtk-svs: improve readability of platform_probe
soc: mediatek: mtk-svs: move svs_platform_probe into probe
soc: mediatek: mtk-svs: delete superfluous platform data entries
Ricardo Ribalda (1):
soc: mediatek: mtk-svs: enable the IRQ later
Roger Lu (8):
soc: mediatek: mtk-svs: restore default voltages when svs_init02()
fail
soc: mediatek: mtk-svs: reset svs when svs_resume() fail
soc: mediatek: mtk-svs: use svs clk control APIs
soc: mediatek: mtk-svs: add thermal voltage compensation if needed
soc: mediatek: mtk-svs: keep svs alive if CONFIG_DEBUG_FS not
supported
soc: mediatek: mtk-svs: use svs get efuse common function
soc: mediatek: mtk-svs: use common function to disable restore
voltages
soc: mtk-svs: mt8183: refactor o_slope calculation
Shang XiaoJing (1):
soc: mediatek: mtk-svs: Use pm_runtime_resume_and_get() in
svs_init01()
drivers/soc/mediatek/mtk-svs.c | 339 +++++++++++++++++----------------
1 file changed, 173 insertions(+), 166 deletions(-)
Some projects might not support CONFIG_DEBUG_FS but still needs svs to be
alive. Therefore, enclose debug cmd codes with CONFIG_DEBUG_FS to make sure
svs can be alive when CONFIG_DEBUG_FS not supported.
Signed-off-by: Roger Lu <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index a3c84e819bc5..70ca9c9acae0 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -138,6 +138,7 @@
static DEFINE_SPINLOCK(svs_lock);
+#ifdef CONFIG_DEBUG_FS
#define debug_fops_ro(name) \
static int svs_##name##_debug_open(struct inode *inode, \
struct file *filp) \
@@ -170,6 +171,7 @@ static DEFINE_SPINLOCK(svs_lock);
}
#define svs_dentry_data(name) {__stringify(name), &svs_##name##_debug_fops}
+#endif
/**
* enum svsb_phase - svs bank phase enumeration
@@ -652,6 +654,7 @@ static int svs_adjust_pm_opp_volts(struct svs_bank *svsb)
return ret;
}
+#ifdef CONFIG_DEBUG_FS
static int svs_dump_debug_show(struct seq_file *m, void *p)
{
struct svs_platform *svsp = (struct svs_platform *)m->private;
@@ -867,6 +870,7 @@ static int svs_create_debug_cmds(struct svs_platform *svsp)
return 0;
}
+#endif /* CONFIG_DEBUG_FS */
static u32 interpolate(u32 f0, u32 f1, u32 v0, u32 v1, u32 fx)
{
@@ -2476,11 +2480,13 @@ static int svs_probe(struct platform_device *pdev)
goto svs_probe_iounmap;
}
+#ifdef CONFIG_DEBUG_FS
ret = svs_create_debug_cmds(svsp);
if (ret) {
dev_err(svsp->dev, "svs create debug cmds fail: %d\n", ret);
goto svs_probe_iounmap;
}
+#endif
return 0;
--
2.18.0
Add svs reset when svs_resume() fail.
Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare() on err in svs_resume()")
Signed-off-by: Roger Lu <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index 2df30a6bca28..a803b92afc3d 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -1614,12 +1614,16 @@ static int svs_resume(struct device *dev)
ret = svs_init02(svsp);
if (ret)
- goto out_of_resume;
+ goto svs_resume_reset_assert;
svs_mon_mode(svsp);
return 0;
+svs_resume_reset_assert:
+ dev_err(svsp->dev, "assert reset: %d\n",
+ reset_control_assert(svsp->rst));
+
out_of_resume:
clk_disable_unprepare(svsp->main_clk);
return ret;
--
2.18.0
From: Matthias Brugger <[email protected]>
We only ever call the SoC specific probe function from
svs_platform_probe. No need to carry that function in a global
datastructure around.
Signed-off-by: Matthias Brugger <[email protected]>
Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
Signed-off-by: Roger Lu <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 5 +----
1 file changed, 1 insertion(+), 4 deletions(-)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index 70ca9c9acae0..41c0ac5ed8d3 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -321,7 +321,6 @@ static const u32 svs_regs_v2[] = {
* @banks: svs banks that svs platform supports
* @rst: svs platform reset control
* @efuse_parsing: svs platform efuse parsing function pointer
- * @probe: svs platform probe function pointer
* @efuse_max: total number of svs efuse
* @tefuse_max: total number of thermal efuse
* @regs: svs platform registers map
@@ -339,7 +338,6 @@ struct svs_platform {
struct svs_bank *banks;
struct reset_control *rst;
bool (*efuse_parsing)(struct svs_platform *svsp);
- int (*probe)(struct svs_platform *svsp);
size_t efuse_max;
size_t tefuse_max;
const u32 *regs;
@@ -2409,11 +2407,10 @@ static struct svs_platform *svs_platform_probe(struct platform_device *pdev)
svsp->name = svsp_data->name;
svsp->banks = svsp_data->banks;
svsp->efuse_parsing = svsp_data->efuse_parsing;
- svsp->probe = svsp_data->probe;
svsp->regs = svsp_data->regs;
svsp->bank_max = svsp_data->bank_max;
- ret = svsp->probe(svsp);
+ ret = svsp_data->probe(svsp);
if (ret)
return ERR_PTR(ret);
--
2.18.0
From: Matthias Brugger <[email protected]>
The platform name and efuse parsing function pointer are only used while
probing the device. Use them from the svs_platform_data struct instead.
Signed-off-by: Matthias Brugger <[email protected]>
Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
Signed-off-by: Roger Lu <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 22 ++++++++++++----------
1 file changed, 12 insertions(+), 10 deletions(-)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index 9d9210c22289..bd23d1111d7b 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -313,14 +313,12 @@ static const u32 svs_regs_v2[] = {
/**
* struct svs_platform - svs platform control
- * @name: svs platform name
* @base: svs platform register base
* @dev: svs platform device
* @main_clk: main clock for svs bank
* @pbank: svs bank pointer needing to be protected by spin_lock section
* @banks: svs banks that svs platform supports
* @rst: svs platform reset control
- * @efuse_parsing: svs platform efuse parsing function pointer
* @efuse_max: total number of svs efuse
* @tefuse_max: total number of thermal efuse
* @regs: svs platform registers map
@@ -330,14 +328,12 @@ static const u32 svs_regs_v2[] = {
* @clk_cnt: clock count shows the clk enable/disable times by svs driver
*/
struct svs_platform {
- char *name;
void __iomem *base;
struct device *dev;
struct clk *main_clk;
struct svs_bank *pbank;
struct svs_bank *banks;
struct reset_control *rst;
- bool (*efuse_parsing)(struct svs_platform *svsp);
size_t efuse_max;
size_t tefuse_max;
const u32 *regs;
@@ -2066,7 +2062,7 @@ static bool svs_is_efuse_data_correct(struct svs_platform *svsp)
svsp->efuse_max /= sizeof(u32);
nvmem_cell_put(cell);
- return svsp->efuse_parsing(svsp);
+ return true;
}
static struct device *svs_get_subsys_device(struct svs_platform *svsp,
@@ -2400,9 +2396,7 @@ static int svs_probe(struct platform_device *pdev)
return -ENOMEM;
svsp->dev = &pdev->dev;
- svsp->name = svsp_data->name;
svsp->banks = svsp_data->banks;
- svsp->efuse_parsing = svsp_data->efuse_parsing;
svsp->regs = svsp_data->regs;
svsp->bank_max = svsp_data->bank_max;
@@ -2413,6 +2407,12 @@ static int svs_probe(struct platform_device *pdev)
if (!svs_is_efuse_data_correct(svsp)) {
dev_notice(svsp->dev, "efuse data isn't correct\n");
ret = -EPERM;
+ goto svs_probe_free_efuse;
+ }
+
+ if (!svsp_data->efuse_parsing(svsp)) {
+ dev_err(svsp->dev, "efuse data parsing failed\n");
+ ret = -EPERM;
goto svs_probe_free_resource;
}
@@ -2448,7 +2448,7 @@ static int svs_probe(struct platform_device *pdev)
}
ret = devm_request_threaded_irq(svsp->dev, svsp_irq, NULL, svs_isr,
- IRQF_ONESHOT, svsp->name, svsp);
+ IRQF_ONESHOT, svsp_data->name, svsp);
if (ret) {
dev_err(svsp->dev, "register irq(%d) failed: %d\n",
svsp_irq, ret);
@@ -2478,11 +2478,13 @@ static int svs_probe(struct platform_device *pdev)
svs_clk_disable(svsp);
svs_probe_free_resource:
- if (!IS_ERR_OR_NULL(svsp->efuse))
- kfree(svsp->efuse);
if (!IS_ERR_OR_NULL(svsp->tefuse))
kfree(svsp->tefuse);
+svs_probe_free_efuse:
+ if (!IS_ERR_OR_NULL(svsp->efuse))
+ kfree(svsp->efuse);
+
return ret;
}
--
2.18.0
In MediaTek HW design, svs and thermal both use the same clk source.
It means that svs clk reference count from CCF includes thermal control
counts. That makes svs driver confuse whether it disabled svs's main clk
or not from CCF's perspective and lead to turn off their shared clk
unexpectedly. Therefore, we add svs clk control APIs to make sure svs's
main clk is controlled well by svs driver itself.
Here is a NG example. Rely on CCF's reference count and cause problem.
thermal probe (clk ref = 1)
-> svs probe (clk ref = 2)
-> svs suspend (clk ref = 1)
-> thermal suspend (clk ref = 0)
-> thermal resume (clk ref = 1)
-> svs resume (encounter error, clk ref = 1)
-> svs suspend (clk ref = 0)
-> thermal suspend (Fail here, thermal HW control w/o clk)
Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare() on err in svs_resume()")
Signed-off-by: Roger Lu <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 57 ++++++++++++++++++++++++++--------
1 file changed, 44 insertions(+), 13 deletions(-)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index 9575aa645643..830263bad81e 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -326,6 +326,7 @@ static const u32 svs_regs_v2[] = {
* @bank_max: total number of svs banks
* @efuse: svs efuse data received from NVMEM framework
* @tefuse: thermal efuse data received from NVMEM framework
+ * @clk_cnt: clock count shows the clk enable/disable times by svs driver
*/
struct svs_platform {
char *name;
@@ -343,6 +344,7 @@ struct svs_platform {
u32 bank_max;
u32 *efuse;
u32 *tefuse;
+ s32 clk_cnt;
};
struct svs_platform_data {
@@ -502,6 +504,32 @@ static void svs_switch_bank(struct svs_platform *svsp)
svs_writel_relaxed(svsp, svsb->core_sel, CORESEL);
}
+static bool svs_is_clk_enabled(struct svs_platform *svsp)
+{
+ return svsp->clk_cnt > 0 ? true : false;
+}
+
+static int svs_clk_enable(struct svs_platform *svsp)
+{
+ int ret;
+
+ ret = clk_prepare_enable(svsp->main_clk);
+ if (ret) {
+ dev_err(svsp->dev, "cannot enable main_clk: %d\n", ret);
+ return ret;
+ }
+
+ svsp->clk_cnt++;
+
+ return 0;
+}
+
+static void svs_clk_disable(struct svs_platform *svsp)
+{
+ clk_disable_unprepare(svsp->main_clk);
+ svsp->clk_cnt--;
+}
+
static u32 svs_bank_volt_to_opp_volt(u32 svsb_volt, u32 svsb_volt_step,
u32 svsb_volt_base)
{
@@ -1569,6 +1597,12 @@ static int svs_suspend(struct device *dev)
int ret;
u32 idx;
+ if (!svs_is_clk_enabled(svsp)) {
+ dev_err(svsp->dev, "svs clk is disabled already (%d)\n",
+ svsp->clk_cnt);
+ return 0;
+ }
+
for (idx = 0; idx < svsp->bank_max; idx++) {
svsb = &svsp->banks[idx];
@@ -1590,7 +1624,7 @@ static int svs_suspend(struct device *dev)
return ret;
}
- clk_disable_unprepare(svsp->main_clk);
+ svs_clk_disable(svsp);
return 0;
}
@@ -1600,16 +1634,14 @@ static int svs_resume(struct device *dev)
struct svs_platform *svsp = dev_get_drvdata(dev);
int ret;
- ret = clk_prepare_enable(svsp->main_clk);
- if (ret) {
- dev_err(svsp->dev, "cannot enable main_clk, disable svs\n");
+ ret = svs_clk_enable(svsp);
+ if (ret)
return ret;
- }
ret = reset_control_deassert(svsp->rst);
if (ret) {
dev_err(svsp->dev, "cannot deassert reset %d\n", ret);
- goto out_of_resume;
+ goto svs_resume_clk_disable;
}
ret = svs_init02(svsp);
@@ -1624,8 +1656,9 @@ static int svs_resume(struct device *dev)
dev_err(svsp->dev, "assert reset: %d\n",
reset_control_assert(svsp->rst));
-out_of_resume:
- clk_disable_unprepare(svsp->main_clk);
+svs_resume_clk_disable:
+ svs_clk_disable(svsp);
+
return ret;
}
@@ -2411,11 +2444,9 @@ static int svs_probe(struct platform_device *pdev)
goto svs_probe_free_resource;
}
- ret = clk_prepare_enable(svsp->main_clk);
- if (ret) {
- dev_err(svsp->dev, "cannot enable main clk: %d\n", ret);
+ ret = svs_clk_enable(svsp);
+ if (ret)
goto svs_probe_free_resource;
- }
svsp->base = of_iomap(svsp->dev->of_node, 0);
if (IS_ERR_OR_NULL(svsp->base)) {
@@ -2456,7 +2487,7 @@ static int svs_probe(struct platform_device *pdev)
iounmap(svsp->base);
svs_probe_clk_disable:
- clk_disable_unprepare(svsp->main_clk);
+ svs_clk_disable(svsp);
svs_probe_free_resource:
if (!IS_ERR_OR_NULL(svsp->efuse))
--
2.18.0
Some extreme test environment may keep IC temperature very low or very high
during system boot stage. For stability concern, we add thermal voltage
compenstation if needed no matter svs bank phase is in init02 or mon mode.
Signed-off-by: Roger Lu <[email protected]>
Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 17 +++++++++--------
1 file changed, 9 insertions(+), 8 deletions(-)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index 830263bad81e..a3c84e819bc5 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -590,7 +590,7 @@ static int svs_adjust_pm_opp_volts(struct svs_bank *svsb)
}
/* Get thermal effect */
- if (svsb->phase == SVSB_PHASE_MON) {
+ if (!IS_ERR_OR_NULL(svsb->tzd)) {
ret = thermal_zone_get_temp(svsb->tzd, &tzone_temp);
if (ret || (svsb->temp > SVSB_TEMP_UPPER_BOUND &&
svsb->temp < SVSB_TEMP_LOWER_BOUND)) {
@@ -605,7 +605,8 @@ static int svs_adjust_pm_opp_volts(struct svs_bank *svsb)
temp_voffset += svsb->tzone_ltemp_voffset;
/* 2-line bank update all opp volts when running mon mode */
- if (svsb->type == SVSB_HIGH || svsb->type == SVSB_LOW) {
+ if (svsb->phase == SVSB_PHASE_MON && (svsb->type == SVSB_HIGH ||
+ svsb->type == SVSB_LOW)) {
opp_start = 0;
opp_stop = svsb->opp_count;
}
@@ -621,11 +622,6 @@ static int svs_adjust_pm_opp_volts(struct svs_bank *svsb)
/* do nothing */
goto unlock_mutex;
case SVSB_PHASE_INIT02:
- svsb_volt = max(svsb->volt[i], svsb->vmin);
- opp_volt = svs_bank_volt_to_opp_volt(svsb_volt,
- svsb->volt_step,
- svsb->volt_base);
- break;
case SVSB_PHASE_MON:
svsb_volt = max(svsb->volt[i] + temp_voffset, svsb->vmin);
opp_volt = svs_bank_volt_to_opp_volt(svsb_volt,
@@ -1728,7 +1724,7 @@ static int svs_bank_resource_setup(struct svs_platform *svsp)
}
}
- if (svsb->mode_support & SVSB_MODE_MON) {
+ if (!IS_ERR_OR_NULL(svsb->tzone_name)) {
svsb->tzd = thermal_zone_get_zone_by_name(svsb->tzone_name);
if (IS_ERR(svsb->tzd)) {
dev_err(svsb->dev, "cannot get \"%s\" thermal zone\n",
@@ -2201,6 +2197,7 @@ static struct svs_bank svs_mt8192_banks[] = {
.type = SVSB_LOW,
.set_freq_pct = svs_set_bank_freq_pct_v3,
.get_volts = svs_get_bank_volts_v3,
+ .tzone_name = "gpu1",
.volt_flags = SVSB_REMOVE_DVTFIXED_VOLT,
.mode_support = SVSB_MODE_INIT02,
.opp_count = MAX_OPP_ENTRIES,
@@ -2218,6 +2215,10 @@ static struct svs_bank svs_mt8192_banks[] = {
.core_sel = 0x0fff0100,
.int_st = BIT(0),
.ctl0 = 0x00540003,
+ .tzone_htemp = 85000,
+ .tzone_htemp_voffset = 0,
+ .tzone_ltemp = 25000,
+ .tzone_ltemp_voffset = 7,
},
{
.sw_id = SVSB_GPU,
--
2.18.0
From: Matthias Brugger <[email protected]>
If a compatible misses a match data entry, then something is wrong in
the development phase, we don't need to check for that at runtime.
Signed-off-by: Matthias Brugger <[email protected]>
Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
Signed-off-by: Roger Lu <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 4 ----
1 file changed, 4 deletions(-)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index 41c0ac5ed8d3..354582825082 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -2394,10 +2394,6 @@ static struct svs_platform *svs_platform_probe(struct platform_device *pdev)
int ret;
svsp_data = of_device_get_match_data(&pdev->dev);
- if (!svsp_data) {
- dev_err(&pdev->dev, "no svs platform data?\n");
- return ERR_PTR(-EPERM);
- }
svsp = devm_kzalloc(&pdev->dev, sizeof(*svsp), GFP_KERNEL);
if (!svsp)
--
2.18.0
The timing of disabling SVS bank and restore default voltage is more
than one place. Therefore, add a common function to use for removing
the superfluous codes.
Signed-off-by: Roger Lu <[email protected]>
Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
---
drivers/soc/mediatek/mtk-svs.c | 54 ++++++++++++++--------------------
1 file changed, 22 insertions(+), 32 deletions(-)
diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
index a7f0a6f02d52..89117807e85d 100644
--- a/drivers/soc/mediatek/mtk-svs.c
+++ b/drivers/soc/mediatek/mtk-svs.c
@@ -648,6 +648,25 @@ static int svs_adjust_pm_opp_volts(struct svs_bank *svsb)
return ret;
}
+static void svs_bank_disable_and_restore_default_volts(struct svs_platform *svsp,
+ struct svs_bank *svsb)
+{
+ unsigned long flags;
+
+ if (svsb->mode_support == SVSB_MODE_ALL_DISABLE)
+ return;
+
+ spin_lock_irqsave(&svs_lock, flags);
+ svsp->pbank = svsb;
+ svs_switch_bank(svsp);
+ svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
+ svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
+ spin_unlock_irqrestore(&svs_lock, flags);
+
+ svsb->phase = SVSB_PHASE_ERROR;
+ svs_adjust_pm_opp_volts(svsb);
+}
+
#ifdef CONFIG_DEBUG_FS
static int svs_dump_debug_show(struct seq_file *m, void *p)
{
@@ -724,7 +743,6 @@ static ssize_t svs_enable_debug_write(struct file *filp,
{
struct svs_bank *svsb = file_inode(filp)->i_private;
struct svs_platform *svsp = dev_get_drvdata(svsb->dev);
- unsigned long flags;
int enabled, ret;
char *buf = NULL;
@@ -740,16 +758,8 @@ static ssize_t svs_enable_debug_write(struct file *filp,
return ret;
if (!enabled) {
- spin_lock_irqsave(&svs_lock, flags);
- svsp->pbank = svsb;
+ svs_bank_disable_and_restore_default_volts(svsp, svsb);
svsb->mode_support = SVSB_MODE_ALL_DISABLE;
- svs_switch_bank(svsp);
- svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
- svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
- spin_unlock_irqrestore(&svs_lock, flags);
-
- svsb->phase = SVSB_PHASE_ERROR;
- svs_adjust_pm_opp_volts(svsb);
}
kfree(buf);
@@ -1532,16 +1542,7 @@ static int svs_init02(struct svs_platform *svsp)
out_of_init02:
for (idx = 0; idx < svsp->bank_max; idx++) {
svsb = &svsp->banks[idx];
-
- spin_lock_irqsave(&svs_lock, flags);
- svsp->pbank = svsb;
- svs_switch_bank(svsp);
- svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
- svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
- spin_unlock_irqrestore(&svs_lock, flags);
-
- svsb->phase = SVSB_PHASE_ERROR;
- svs_adjust_pm_opp_volts(svsb);
+ svs_bank_disable_and_restore_default_volts(svsp, svsb);
}
return ret;
@@ -1587,7 +1588,6 @@ static int svs_suspend(struct device *dev)
{
struct svs_platform *svsp = dev_get_drvdata(dev);
struct svs_bank *svsb;
- unsigned long flags;
int ret;
u32 idx;
@@ -1599,17 +1599,7 @@ static int svs_suspend(struct device *dev)
for (idx = 0; idx < svsp->bank_max; idx++) {
svsb = &svsp->banks[idx];
-
- /* This might wait for svs_isr() process */
- spin_lock_irqsave(&svs_lock, flags);
- svsp->pbank = svsb;
- svs_switch_bank(svsp);
- svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
- svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
- spin_unlock_irqrestore(&svs_lock, flags);
-
- svsb->phase = SVSB_PHASE_ERROR;
- svs_adjust_pm_opp_volts(svsb);
+ svs_bank_disable_and_restore_default_volts(svsp, svsb);
}
ret = reset_control_assert(svsp->rst);
--
2.18.0
On 11/01/2023 08:45, Roger Lu wrote:
> In MediaTek HW design, svs and thermal both use the same clk source.
> It means that svs clk reference count from CCF includes thermal control
> counts. That makes svs driver confuse whether it disabled svs's main clk
> or not from CCF's perspective and lead to turn off their shared clk
> unexpectedly. Therefore, we add svs clk control APIs to make sure svs's
> main clk is controlled well by svs driver itself.
>
> Here is a NG example. Rely on CCF's reference count and cause problem.
>
> thermal probe (clk ref = 1)
> -> svs probe (clk ref = 2)
> -> svs suspend (clk ref = 1)
> -> thermal suspend (clk ref = 0)
> -> thermal resume (clk ref = 1)
> -> svs resume (encounter error, clk ref = 1)
> -> svs suspend (clk ref = 0)
> -> thermal suspend (Fail here, thermal HW control w/o clk)
>
> Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare() on err in svs_resume()")
> Signed-off-by: Roger Lu <[email protected]>
That looks wrong. Although I don't out of my mind, there should be a way to tell
the clock framework that this clock is shared between several devices.
I wonder if using clk_enable and clk_disable in svs_resume/suspend wouldn't be
enough.
Regards,
Matthias
> ---
> drivers/soc/mediatek/mtk-svs.c | 57 ++++++++++++++++++++++++++--------
> 1 file changed, 44 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
> index 9575aa645643..830263bad81e 100644
> --- a/drivers/soc/mediatek/mtk-svs.c
> +++ b/drivers/soc/mediatek/mtk-svs.c
> @@ -326,6 +326,7 @@ static const u32 svs_regs_v2[] = {
> * @bank_max: total number of svs banks
> * @efuse: svs efuse data received from NVMEM framework
> * @tefuse: thermal efuse data received from NVMEM framework
> + * @clk_cnt: clock count shows the clk enable/disable times by svs driver
> */
> struct svs_platform {
> char *name;
> @@ -343,6 +344,7 @@ struct svs_platform {
> u32 bank_max;
> u32 *efuse;
> u32 *tefuse;
> + s32 clk_cnt;
> };
>
> struct svs_platform_data {
> @@ -502,6 +504,32 @@ static void svs_switch_bank(struct svs_platform *svsp)
> svs_writel_relaxed(svsp, svsb->core_sel, CORESEL);
> }
>
> +static bool svs_is_clk_enabled(struct svs_platform *svsp)
> +{
> + return svsp->clk_cnt > 0 ? true : false;
> +}
> +
> +static int svs_clk_enable(struct svs_platform *svsp)
> +{
> + int ret;
> +
> + ret = clk_prepare_enable(svsp->main_clk);
> + if (ret) {
> + dev_err(svsp->dev, "cannot enable main_clk: %d\n", ret);
> + return ret;
> + }
> +
> + svsp->clk_cnt++;
> +
> + return 0;
> +}
> +
> +static void svs_clk_disable(struct svs_platform *svsp)
> +{
> + clk_disable_unprepare(svsp->main_clk);
> + svsp->clk_cnt--;
> +}
> +
> static u32 svs_bank_volt_to_opp_volt(u32 svsb_volt, u32 svsb_volt_step,
> u32 svsb_volt_base)
> {
> @@ -1569,6 +1597,12 @@ static int svs_suspend(struct device *dev)
> int ret;
> u32 idx;
>
> + if (!svs_is_clk_enabled(svsp)) {
> + dev_err(svsp->dev, "svs clk is disabled already (%d)\n",
> + svsp->clk_cnt);
> + return 0;
> + }
> +
> for (idx = 0; idx < svsp->bank_max; idx++) {
> svsb = &svsp->banks[idx];
>
> @@ -1590,7 +1624,7 @@ static int svs_suspend(struct device *dev)
> return ret;
> }
>
> - clk_disable_unprepare(svsp->main_clk);
> + svs_clk_disable(svsp);
>
> return 0;
> }
> @@ -1600,16 +1634,14 @@ static int svs_resume(struct device *dev)
> struct svs_platform *svsp = dev_get_drvdata(dev);
> int ret;
>
> - ret = clk_prepare_enable(svsp->main_clk);
> - if (ret) {
> - dev_err(svsp->dev, "cannot enable main_clk, disable svs\n");
> + ret = svs_clk_enable(svsp);
> + if (ret)
> return ret;
> - }
>
> ret = reset_control_deassert(svsp->rst);
> if (ret) {
> dev_err(svsp->dev, "cannot deassert reset %d\n", ret);
> - goto out_of_resume;
> + goto svs_resume_clk_disable;
> }
>
> ret = svs_init02(svsp);
> @@ -1624,8 +1656,9 @@ static int svs_resume(struct device *dev)
> dev_err(svsp->dev, "assert reset: %d\n",
> reset_control_assert(svsp->rst));
>
> -out_of_resume:
> - clk_disable_unprepare(svsp->main_clk);
> +svs_resume_clk_disable:
> + svs_clk_disable(svsp);
> +
> return ret;
> }
>
> @@ -2411,11 +2444,9 @@ static int svs_probe(struct platform_device *pdev)
> goto svs_probe_free_resource;
> }
>
> - ret = clk_prepare_enable(svsp->main_clk);
> - if (ret) {
> - dev_err(svsp->dev, "cannot enable main clk: %d\n", ret);
> + ret = svs_clk_enable(svsp);
> + if (ret)
> goto svs_probe_free_resource;
> - }
>
> svsp->base = of_iomap(svsp->dev->of_node, 0);
> if (IS_ERR_OR_NULL(svsp->base)) {
> @@ -2456,7 +2487,7 @@ static int svs_probe(struct platform_device *pdev)
> iounmap(svsp->base);
>
> svs_probe_clk_disable:
> - clk_disable_unprepare(svsp->main_clk);
> + svs_clk_disable(svsp);
>
> svs_probe_free_resource:
> if (!IS_ERR_OR_NULL(svsp->efuse))
On 11/01/2023 08:45, Roger Lu wrote:
> Add svs reset when svs_resume() fail.
>
> Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare() on err in svs_resume()")
> Signed-off-by: Roger Lu <[email protected]>
Applied, thanks!
> ---
> drivers/soc/mediatek/mtk-svs.c | 6 +++++-
> 1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
> index 2df30a6bca28..a803b92afc3d 100644
> --- a/drivers/soc/mediatek/mtk-svs.c
> +++ b/drivers/soc/mediatek/mtk-svs.c
> @@ -1614,12 +1614,16 @@ static int svs_resume(struct device *dev)
>
> ret = svs_init02(svsp);
> if (ret)
> - goto out_of_resume;
> + goto svs_resume_reset_assert;
>
> svs_mon_mode(svsp);
>
> return 0;
>
> +svs_resume_reset_assert:
> + dev_err(svsp->dev, "assert reset: %d\n",
> + reset_control_assert(svsp->rst));
> +
> out_of_resume:
> clk_disable_unprepare(svsp->main_clk);
> return ret;
On 11/01/2023 08:45, Roger Lu wrote:
> Some projects might not support CONFIG_DEBUG_FS but still needs svs to be
> alive. Therefore, enclose debug cmd codes with CONFIG_DEBUG_FS to make sure
> svs can be alive when CONFIG_DEBUG_FS not supported.
>
> Signed-off-by: Roger Lu <[email protected]>
Applied, thanks!
> ---
> drivers/soc/mediatek/mtk-svs.c | 6 ++++++
> 1 file changed, 6 insertions(+)
>
> diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
> index a3c84e819bc5..70ca9c9acae0 100644
> --- a/drivers/soc/mediatek/mtk-svs.c
> +++ b/drivers/soc/mediatek/mtk-svs.c
> @@ -138,6 +138,7 @@
>
> static DEFINE_SPINLOCK(svs_lock);
>
> +#ifdef CONFIG_DEBUG_FS
> #define debug_fops_ro(name) \
> static int svs_##name##_debug_open(struct inode *inode, \
> struct file *filp) \
> @@ -170,6 +171,7 @@ static DEFINE_SPINLOCK(svs_lock);
> }
>
> #define svs_dentry_data(name) {__stringify(name), &svs_##name##_debug_fops}
> +#endif
>
> /**
> * enum svsb_phase - svs bank phase enumeration
> @@ -652,6 +654,7 @@ static int svs_adjust_pm_opp_volts(struct svs_bank *svsb)
> return ret;
> }
>
> +#ifdef CONFIG_DEBUG_FS
> static int svs_dump_debug_show(struct seq_file *m, void *p)
> {
> struct svs_platform *svsp = (struct svs_platform *)m->private;
> @@ -867,6 +870,7 @@ static int svs_create_debug_cmds(struct svs_platform *svsp)
>
> return 0;
> }
> +#endif /* CONFIG_DEBUG_FS */
>
> static u32 interpolate(u32 f0, u32 f1, u32 v0, u32 v1, u32 fx)
> {
> @@ -2476,11 +2480,13 @@ static int svs_probe(struct platform_device *pdev)
> goto svs_probe_iounmap;
> }
>
> +#ifdef CONFIG_DEBUG_FS
> ret = svs_create_debug_cmds(svsp);
> if (ret) {
> dev_err(svsp->dev, "svs create debug cmds fail: %d\n", ret);
> goto svs_probe_iounmap;
> }
> +#endif
>
> return 0;
>
On 11/01/2023 08:45, Roger Lu wrote:
> From: Matthias Brugger <[email protected]>
>
> We only ever call the SoC specific probe function from
> svs_platform_probe. No need to carry that function in a global
> datastructure around.
>
> Signed-off-by: Matthias Brugger <[email protected]>
> Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
> Signed-off-by: Roger Lu <[email protected]>
Applied, thanks!
> ---
> drivers/soc/mediatek/mtk-svs.c | 5 +----
> 1 file changed, 1 insertion(+), 4 deletions(-)
>
> diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
> index 70ca9c9acae0..41c0ac5ed8d3 100644
> --- a/drivers/soc/mediatek/mtk-svs.c
> +++ b/drivers/soc/mediatek/mtk-svs.c
> @@ -321,7 +321,6 @@ static const u32 svs_regs_v2[] = {
> * @banks: svs banks that svs platform supports
> * @rst: svs platform reset control
> * @efuse_parsing: svs platform efuse parsing function pointer
> - * @probe: svs platform probe function pointer
> * @efuse_max: total number of svs efuse
> * @tefuse_max: total number of thermal efuse
> * @regs: svs platform registers map
> @@ -339,7 +338,6 @@ struct svs_platform {
> struct svs_bank *banks;
> struct reset_control *rst;
> bool (*efuse_parsing)(struct svs_platform *svsp);
> - int (*probe)(struct svs_platform *svsp);
> size_t efuse_max;
> size_t tefuse_max;
> const u32 *regs;
> @@ -2409,11 +2407,10 @@ static struct svs_platform *svs_platform_probe(struct platform_device *pdev)
> svsp->name = svsp_data->name;
> svsp->banks = svsp_data->banks;
> svsp->efuse_parsing = svsp_data->efuse_parsing;
> - svsp->probe = svsp_data->probe;
> svsp->regs = svsp_data->regs;
> svsp->bank_max = svsp_data->bank_max;
>
> - ret = svsp->probe(svsp);
> + ret = svsp_data->probe(svsp);
> if (ret)
> return ERR_PTR(ret);
>
On 11/01/2023 08:45, Roger Lu wrote:
> From: Matthias Brugger <[email protected]>
>
> If a compatible misses a match data entry, then something is wrong in
> the development phase, we don't need to check for that at runtime.
>
> Signed-off-by: Matthias Brugger <[email protected]>
> Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
> Signed-off-by: Roger Lu <[email protected]>
Applied, thanks!
> ---
> drivers/soc/mediatek/mtk-svs.c | 4 ----
> 1 file changed, 4 deletions(-)
>
> diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
> index 41c0ac5ed8d3..354582825082 100644
> --- a/drivers/soc/mediatek/mtk-svs.c
> +++ b/drivers/soc/mediatek/mtk-svs.c
> @@ -2394,10 +2394,6 @@ static struct svs_platform *svs_platform_probe(struct platform_device *pdev)
> int ret;
>
> svsp_data = of_device_get_match_data(&pdev->dev);
> - if (!svsp_data) {
> - dev_err(&pdev->dev, "no svs platform data?\n");
> - return ERR_PTR(-EPERM);
> - }
>
> svsp = devm_kzalloc(&pdev->dev, sizeof(*svsp), GFP_KERNEL);
> if (!svsp)
On 11/01/2023 08:45, Roger Lu wrote:
> From: Matthias Brugger <[email protected]>
>
> The platform name and efuse parsing function pointer are only used while
> probing the device. Use them from the svs_platform_data struct instead.
>
> Signed-off-by: Matthias Brugger <[email protected]>
> Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
> Signed-off-by: Roger Lu <[email protected]>
Applied, thanks!
> ---
> drivers/soc/mediatek/mtk-svs.c | 22 ++++++++++++----------
> 1 file changed, 12 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
> index 9d9210c22289..bd23d1111d7b 100644
> --- a/drivers/soc/mediatek/mtk-svs.c
> +++ b/drivers/soc/mediatek/mtk-svs.c
> @@ -313,14 +313,12 @@ static const u32 svs_regs_v2[] = {
>
> /**
> * struct svs_platform - svs platform control
> - * @name: svs platform name
> * @base: svs platform register base
> * @dev: svs platform device
> * @main_clk: main clock for svs bank
> * @pbank: svs bank pointer needing to be protected by spin_lock section
> * @banks: svs banks that svs platform supports
> * @rst: svs platform reset control
> - * @efuse_parsing: svs platform efuse parsing function pointer
> * @efuse_max: total number of svs efuse
> * @tefuse_max: total number of thermal efuse
> * @regs: svs platform registers map
> @@ -330,14 +328,12 @@ static const u32 svs_regs_v2[] = {
> * @clk_cnt: clock count shows the clk enable/disable times by svs driver
> */
> struct svs_platform {
> - char *name;
> void __iomem *base;
> struct device *dev;
> struct clk *main_clk;
> struct svs_bank *pbank;
> struct svs_bank *banks;
> struct reset_control *rst;
> - bool (*efuse_parsing)(struct svs_platform *svsp);
> size_t efuse_max;
> size_t tefuse_max;
> const u32 *regs;
> @@ -2066,7 +2062,7 @@ static bool svs_is_efuse_data_correct(struct svs_platform *svsp)
> svsp->efuse_max /= sizeof(u32);
> nvmem_cell_put(cell);
>
> - return svsp->efuse_parsing(svsp);
> + return true;
> }
>
> static struct device *svs_get_subsys_device(struct svs_platform *svsp,
> @@ -2400,9 +2396,7 @@ static int svs_probe(struct platform_device *pdev)
> return -ENOMEM;
>
> svsp->dev = &pdev->dev;
> - svsp->name = svsp_data->name;
> svsp->banks = svsp_data->banks;
> - svsp->efuse_parsing = svsp_data->efuse_parsing;
> svsp->regs = svsp_data->regs;
> svsp->bank_max = svsp_data->bank_max;
>
> @@ -2413,6 +2407,12 @@ static int svs_probe(struct platform_device *pdev)
> if (!svs_is_efuse_data_correct(svsp)) {
> dev_notice(svsp->dev, "efuse data isn't correct\n");
> ret = -EPERM;
> + goto svs_probe_free_efuse;
> + }
> +
> + if (!svsp_data->efuse_parsing(svsp)) {
> + dev_err(svsp->dev, "efuse data parsing failed\n");
> + ret = -EPERM;
> goto svs_probe_free_resource;
> }
>
> @@ -2448,7 +2448,7 @@ static int svs_probe(struct platform_device *pdev)
> }
>
> ret = devm_request_threaded_irq(svsp->dev, svsp_irq, NULL, svs_isr,
> - IRQF_ONESHOT, svsp->name, svsp);
> + IRQF_ONESHOT, svsp_data->name, svsp);
> if (ret) {
> dev_err(svsp->dev, "register irq(%d) failed: %d\n",
> svsp_irq, ret);
> @@ -2478,11 +2478,13 @@ static int svs_probe(struct platform_device *pdev)
> svs_clk_disable(svsp);
>
> svs_probe_free_resource:
> - if (!IS_ERR_OR_NULL(svsp->efuse))
> - kfree(svsp->efuse);
> if (!IS_ERR_OR_NULL(svsp->tefuse))
> kfree(svsp->tefuse);
>
> +svs_probe_free_efuse:
> + if (!IS_ERR_OR_NULL(svsp->efuse))
> + kfree(svsp->efuse);
> +
> return ret;
> }
>
On 11/01/2023 08:45, Roger Lu wrote:
> The timing of disabling SVS bank and restore default voltage is more
> than one place. Therefore, add a common function to use for removing
> the superfluous codes.
>
> Signed-off-by: Roger Lu <[email protected]>
> Reviewed-by: AngeloGioacchino Del Regno <[email protected]>
Same here, change looks good. Could you please rebase and resend:
Reviewed-by: Matthias Brugger <[email protected]>
> ---
> drivers/soc/mediatek/mtk-svs.c | 54 ++++++++++++++--------------------
> 1 file changed, 22 insertions(+), 32 deletions(-)
>
> diff --git a/drivers/soc/mediatek/mtk-svs.c b/drivers/soc/mediatek/mtk-svs.c
> index a7f0a6f02d52..89117807e85d 100644
> --- a/drivers/soc/mediatek/mtk-svs.c
> +++ b/drivers/soc/mediatek/mtk-svs.c
> @@ -648,6 +648,25 @@ static int svs_adjust_pm_opp_volts(struct svs_bank *svsb)
> return ret;
> }
>
> +static void svs_bank_disable_and_restore_default_volts(struct svs_platform *svsp,
> + struct svs_bank *svsb)
> +{
> + unsigned long flags;
> +
> + if (svsb->mode_support == SVSB_MODE_ALL_DISABLE)
> + return;
> +
> + spin_lock_irqsave(&svs_lock, flags);
> + svsp->pbank = svsb;
> + svs_switch_bank(svsp);
> + svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
> + svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
> + spin_unlock_irqrestore(&svs_lock, flags);
> +
> + svsb->phase = SVSB_PHASE_ERROR;
> + svs_adjust_pm_opp_volts(svsb);
> +}
> +
> #ifdef CONFIG_DEBUG_FS
> static int svs_dump_debug_show(struct seq_file *m, void *p)
> {
> @@ -724,7 +743,6 @@ static ssize_t svs_enable_debug_write(struct file *filp,
> {
> struct svs_bank *svsb = file_inode(filp)->i_private;
> struct svs_platform *svsp = dev_get_drvdata(svsb->dev);
> - unsigned long flags;
> int enabled, ret;
> char *buf = NULL;
>
> @@ -740,16 +758,8 @@ static ssize_t svs_enable_debug_write(struct file *filp,
> return ret;
>
> if (!enabled) {
> - spin_lock_irqsave(&svs_lock, flags);
> - svsp->pbank = svsb;
> + svs_bank_disable_and_restore_default_volts(svsp, svsb);
> svsb->mode_support = SVSB_MODE_ALL_DISABLE;
> - svs_switch_bank(svsp);
> - svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
> - svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
> - spin_unlock_irqrestore(&svs_lock, flags);
> -
> - svsb->phase = SVSB_PHASE_ERROR;
> - svs_adjust_pm_opp_volts(svsb);
> }
>
> kfree(buf);
> @@ -1532,16 +1542,7 @@ static int svs_init02(struct svs_platform *svsp)
> out_of_init02:
> for (idx = 0; idx < svsp->bank_max; idx++) {
> svsb = &svsp->banks[idx];
> -
> - spin_lock_irqsave(&svs_lock, flags);
> - svsp->pbank = svsb;
> - svs_switch_bank(svsp);
> - svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
> - svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
> - spin_unlock_irqrestore(&svs_lock, flags);
> -
> - svsb->phase = SVSB_PHASE_ERROR;
> - svs_adjust_pm_opp_volts(svsb);
> + svs_bank_disable_and_restore_default_volts(svsp, svsb);
> }
>
> return ret;
> @@ -1587,7 +1588,6 @@ static int svs_suspend(struct device *dev)
> {
> struct svs_platform *svsp = dev_get_drvdata(dev);
> struct svs_bank *svsb;
> - unsigned long flags;
> int ret;
> u32 idx;
>
> @@ -1599,17 +1599,7 @@ static int svs_suspend(struct device *dev)
>
> for (idx = 0; idx < svsp->bank_max; idx++) {
> svsb = &svsp->banks[idx];
> -
> - /* This might wait for svs_isr() process */
> - spin_lock_irqsave(&svs_lock, flags);
> - svsp->pbank = svsb;
> - svs_switch_bank(svsp);
> - svs_writel_relaxed(svsp, SVSB_PTPEN_OFF, SVSEN);
> - svs_writel_relaxed(svsp, SVSB_INTSTS_VAL_CLEAN, INTSTS);
> - spin_unlock_irqrestore(&svs_lock, flags);
> -
> - svsb->phase = SVSB_PHASE_ERROR;
> - svs_adjust_pm_opp_volts(svsb);
> + svs_bank_disable_and_restore_default_volts(svsp, svsb);
> }
>
> ret = reset_control_assert(svsp->rst);
Hi Matthias Sir,
On Tue, 2023-01-31 at 14:40 +0100, Matthias Brugger wrote:
>
> On 11/01/2023 08:45, Roger Lu wrote:
> > The timing of disabling SVS bank and restore default voltage is more
> > than one place. Therefore, add a common function to use for removing
> > the superfluous codes.
> >
> > Signed-off-by: Roger Lu <[email protected]>
> > Reviewed-by: AngeloGioacchino Del Regno <
> > [email protected]>
>
> Same here, change looks good. Could you please rebase and resend:
No Problem, I'll rebase and resend it. Thanks.
>
> Reviewed-by: Matthias Brugger <[email protected]>
... [snip] ...
Hi Matthias Sir,
On Tue, 2023-01-31 at 14:19 +0100, Matthias Brugger wrote:
>
> On 11/01/2023 08:45, Roger Lu wrote:
> > In MediaTek HW design, svs and thermal both use the same clk source.
> > It means that svs clk reference count from CCF includes thermal control
> > counts. That makes svs driver confuse whether it disabled svs's main clk
> > or not from CCF's perspective and lead to turn off their shared clk
> > unexpectedly. Therefore, we add svs clk control APIs to make sure svs's
> > main clk is controlled well by svs driver itself.
> >
> > Here is a NG example. Rely on CCF's reference count and cause problem.
> >
> > thermal probe (clk ref = 1)
> > -> svs probe (clk ref = 2)
> > -> svs suspend (clk ref = 1)
> > -> thermal suspend (clk ref = 0)
> > -> thermal resume (clk ref = 1)
> > -> svs resume (encounter error, clk ref = 1)
> > -> svs suspend (clk ref = 0)
> > -> thermal suspend (Fail here, thermal HW control w/o clk)
> >
> > Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare() on
> > err in svs_resume()")
> > Signed-off-by: Roger Lu <[email protected]>
>
> That looks wrong. Although I don't out of my mind, there should be a way to
> tell
> the clock framework that this clock is shared between several devices.
>
> I wonder if using clk_enable and clk_disable in svs_resume/suspend wouldn't
> be
> enough.
Oh yes, Common Clock Framework (CCF) knows the clock shared between several
devices and maintains clock "on/off" by reference count.
We concern how to stop running svs_suspend() when svs clk is already disabled by
svs_resume(). Take an example as below, if we refers to __clk_is_enabled()
result for knowing svs clk status, it will return "true" all the time because
thermal clk is still on. This causes the problem mentioned in commit message.
static int svs_suspend(struct device *dev)
{
...
if (!__clk_is_enabled(svsp->main_clk)) //always get `true`
return 0;
...
}
>
> Regards,
> Matthias
... [snip] ...
你好，
On 01/02/2023 13:28, Roger Lu (陸瑞傑) wrote:
> Hi Matthias Sir,
>
> On Tue, 2023-01-31 at 14:19 +0100, Matthias Brugger wrote:
>>
>> On 11/01/2023 08:45, Roger Lu wrote:
>>> In MediaTek HW design, svs and thermal both use the same clk source.
>>> It means that svs clk reference count from CCF includes thermal control
>>> counts. That makes svs driver confuse whether it disabled svs's main clk
>>> or not from CCF's perspective and lead to turn off their shared clk
>>> unexpectedly. Therefore, we add svs clk control APIs to make sure svs's
>>> main clk is controlled well by svs driver itself.
>>>
>>> Here is a NG example. Rely on CCF's reference count and cause problem.
>>>
>>> thermal probe (clk ref = 1)
>>> -> svs probe (clk ref = 2)
>>> -> svs suspend (clk ref = 1)
>>> -> thermal suspend (clk ref = 0)
>>> -> thermal resume (clk ref = 1)
>>> -> svs resume (encounter error, clk ref = 1)
>>> -> svs suspend (clk ref = 0)
>>> -> thermal suspend (Fail here, thermal HW control w/o clk)
>>>
>>> Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare() on
>>> err in svs_resume()")
>>> Signed-off-by: Roger Lu <[email protected]>
>>
>> That looks wrong. Although I don't out of my mind, there should be a way to
>> tell
>> the clock framework that this clock is shared between several devices.
>>
>> I wonder if using clk_enable and clk_disable in svs_resume/suspend wouldn't
>> be
>> enough.
>
> Oh yes, Common Clock Framework (CCF) knows the clock shared between several
> devices and maintains clock "on/off" by reference count.
>
The thing is if you use clk_prepare_enable then the clock framework check's if
the clock is already prepared, which could happen like you described in the
svs_resume (encount error) case in the commit message. The question is, can't we
just use clk_enable and clk_disable in resume/suspend and only prepare the clock
in the probe function?
> We concern how to stop running svs_suspend() when svs clk is already disabled by
> svs_resume(). Take an example as below, if we refers to __clk_is_enabled()
> result for knowing svs clk status, it will return "true" all the time because
> thermal clk is still on. This causes the problem mentioned in commit message.
>
I would expect that the kernel takes care that we can't enter a resume path for
a device before the suspend path has finished. Honestly I don't really
understand the problem here. It seems something different then what you
described in the commit message.
Please help me understand better.
谢谢，再见
Matthias
> static int svs_suspend(struct device *dev)
> {
> ...
> if (!__clk_is_enabled(svsp->main_clk)) //always get `true`
> return 0;
> ...
> }
>
>>
>> Regards,
>> Matthias
>
> ... [snip] ...
Hi Matthias Sir,
On Thu, 2023-02-02 at 11:29 +0100, Matthias Brugger wrote:
> 你好，
I got shock and thought someone used your name to reply. However,
your email account helps me clear my mind. Haha.. Nice and warm to see mandarin
on patchwork. It's so fresh and exciting :-).
>
> On 01/02/2023 13:28, Roger Lu (陸瑞傑) wrote:
> > Hi Matthias Sir,
> >
> > On Tue, 2023-01-31 at 14:19 +0100, Matthias Brugger wrote:
> > >
> > > On 11/01/2023 08:45, Roger Lu wrote:
> > > > In MediaTek HW design, svs and thermal both use the same clk source.
> > > > It means that svs clk reference count from CCF includes thermal control
> > > > counts. That makes svs driver confuse whether it disabled svs's main clk
> > > > or not from CCF's perspective and lead to turn off their shared clk
> > > > unexpectedly. Therefore, we add svs clk control APIs to make sure svs's
> > > > main clk is controlled well by svs driver itself.
> > > >
> > > > Here is a NG example. Rely on CCF's reference count and cause problem.
> > > >
> > > > thermal probe (clk ref = 1)
> > > > -> svs probe (clk ref = 2)
> > > > -> svs suspend (clk ref = 1)
> > > > -> thermal suspend (clk ref = 0)
> > > > -> thermal resume (clk ref = 1)
> > > > -> svs resume (encounter error, clk ref = 1)
> > > > -> svs suspend (clk ref = 0)
> > > > -> thermal suspend (Fail here, thermal HW control w/o clk)
> > > >
> > > > Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare()
> > > > on
> > > > err in svs_resume()")
> > > > Signed-off-by: Roger Lu <[email protected]>
> > >
> > > That looks wrong. Although I don't out of my mind, there should be a way
> > > to
> > > tell
> > > the clock framework that this clock is shared between several devices.
> > >
> > > I wonder if using clk_enable and clk_disable in svs_resume/suspend
> > > wouldn't
> > > be
> > > enough.
> >
> > Oh yes, Common Clock Framework (CCF) knows the clock shared between several
> > devices and maintains clock "on/off" by reference count.
> >
>
> The thing is if you use clk_prepare_enable then the clock framework check's
> if
> the clock is already prepared, which could happen like you described in the
> svs_resume (encount error) case in the commit message. The question is, can't
> we
> just use clk_enable and clk_disable in resume/suspend and only prepare the
> clock
> in the probe function?
We'll think if this can fix the problem. Thanks for the advice very much.
>
> > We concern how to stop running svs_suspend() when svs clk is already
> > disabled by
> > svs_resume(). Take an example as below, if we refers to __clk_is_enabled()
> > result for knowing svs clk status, it will return "true" all the time
> > because
> > thermal clk is still on. This causes the problem mentioned in commit
> > message.
> >
>
> I would expect that the kernel takes care that we can't enter a resume path
> for
> a device before the suspend path has finished. Honestly I don't really
> understand the problem here. It seems something different then what you
> described in the commit message.
>
> Please help me understand better.
I see. This patch title needs to be changed to "avoid turning off svs clk twice
unexpectedly" for pointing out the problem precisely. We saw a loophole that svs
clk might be turned off in svs_resume() first and in svs_suspend() again without
enabling svs clk during these the process. Therefore, we try to fix it by this
patch. Below is our thinking process to explain how we got here.
1. (abandoned) We add __clk_is_enabled() check in svs_suspend() to prevent svs
clk from being turned off twice when svs_resume() turned off svs clk in the
error-handling process. Nonetheless, we met the NG case in the commit message.
2. (current patch) We add svs clk control hint to understand if we need to run
svs_suspend() or not if svs_resume() turned off svs clk before.
>
> 谢谢，再见
:-)
Sincerely,
Roger Lu
On 06/02/2023 03:01, Roger Lu (陸瑞傑) wrote:
> Hi Matthias Sir,
>
>
> On Thu, 2023-02-02 at 11:29 +0100, Matthias Brugger wrote:
>> 你好，
>
> I got shock and thought someone used your name to reply. However,
> your email account helps me clear my mind. Haha.. Nice and warm to see mandarin
> on patchwork. It's so fresh and exciting :-).
>
谢谢。 I'm learning mainland Chinese for a few month now, I also learned that you
use different symbols in Taiwan, which I don't know. 对不起。
>>
>> On 01/02/2023 13:28, Roger Lu (陸瑞傑) wrote:
>>> Hi Matthias Sir,
>>>
>>> On Tue, 2023-01-31 at 14:19 +0100, Matthias Brugger wrote:
>>>>
>>>> On 11/01/2023 08:45, Roger Lu wrote:
>>>>> In MediaTek HW design, svs and thermal both use the same clk source.
>>>>> It means that svs clk reference count from CCF includes thermal control
>>>>> counts. That makes svs driver confuse whether it disabled svs's main clk
>>>>> or not from CCF's perspective and lead to turn off their shared clk
>>>>> unexpectedly. Therefore, we add svs clk control APIs to make sure svs's
>>>>> main clk is controlled well by svs driver itself.
>>>>>
>>>>> Here is a NG example. Rely on CCF's reference count and cause problem.
>>>>>
>>>>> thermal probe (clk ref = 1)
>>>>> -> svs probe (clk ref = 2)
>>>>> -> svs suspend (clk ref = 1)
>>>>> -> thermal suspend (clk ref = 0)
>>>>> -> thermal resume (clk ref = 1)
>>>>> -> svs resume (encounter error, clk ref = 1)
>>>>> -> svs suspend (clk ref = 0)
>>>>> -> thermal suspend (Fail here, thermal HW control w/o clk)
>>>>>
>>>>> Fixes: a825d72f74a3 ("soc: mediatek: fix missing clk_disable_unprepare()
>>>>> on
>>>>> err in svs_resume()")
>>>>> Signed-off-by: Roger Lu <[email protected]>
>>>>
>>>> That looks wrong. Although I don't out of my mind, there should be a way
>>>> to
>>>> tell
>>>> the clock framework that this clock is shared between several devices.
>>>>
>>>> I wonder if using clk_enable and clk_disable in svs_resume/suspend
>>>> wouldn't
>>>> be
>>>> enough.
>>>
>>> Oh yes, Common Clock Framework (CCF) knows the clock shared between several
>>> devices and maintains clock "on/off" by reference count.
>>>
>>
>> The thing is if you use clk_prepare_enable then the clock framework check's
>> if
>> the clock is already prepared, which could happen like you described in the
>> svs_resume (encount error) case in the commit message. The question is, can't
>> we
>> just use clk_enable and clk_disable in resume/suspend and only prepare the
>> clock
>> in the probe function?
>
> We'll think if this can fix the problem. Thanks for the advice very much.
>
>>
>>> We concern how to stop running svs_suspend() when svs clk is already
>>> disabled by
>>> svs_resume(). Take an example as below, if we refers to __clk_is_enabled()
>>> result for knowing svs clk status, it will return "true" all the time
>>> because
>>> thermal clk is still on. This causes the problem mentioned in commit
>>> message.
>>>
>>
>> I would expect that the kernel takes care that we can't enter a resume path
>> for
>> a device before the suspend path has finished. Honestly I don't really
>> understand the problem here. It seems something different then what you
>> described in the commit message.
>>
>> Please help me understand better.
>
> I see. This patch title needs to be changed to "avoid turning off svs clk twice
> unexpectedly" for pointing out the problem precisely. We saw a loophole that svs
> clk might be turned off in svs_resume() first and in svs_suspend() again without
> enabling svs clk during these the process. Therefore, we try to fix it by this
> patch. Below is our thinking process to explain how we got here.
>
> 1. (abandoned) We add __clk_is_enabled() check in svs_suspend() to prevent svs
> clk from being turned off twice when svs_resume() turned off svs clk in the
> error-handling process. Nonetheless, we met the NG case in the commit message.
> 2. (current patch) We add svs clk control hint to understand if we need to run
> svs_suspend() or not if svs_resume() turned off svs clk before.
>
Did you had a look on the dev_pm_ops？ Maybe we can use suspend_late,
resume_early to make sure there is no race condition. I wonder also if we can't
make sure that this does not happen using device links. Sorry, I can't give
better guidance on how to use this technologies, but I have the feeling we can
fix this with existing infrastructure.
再见。
Matthias
>>
>> 谢谢，再见
>
> :-)
>
>
> Sincerely,
> Roger Lu
Hi Matthias Sir,
On Mon, 2023-02-06 at 13:09 +0100, Matthias Brugger wrote:
>
> On 06/02/2023 03:01, Roger Lu (陸瑞傑) wrote:
> > Hi Matthias Sir,
> >
> >
> > On Thu, 2023-02-02 at 11:29 +0100, Matthias Brugger wrote:
> > > 你好，
> >
> > I got shock and thought someone used your name to reply. However,
> > your email account helps me clear my mind. Haha.. Nice and warm to see
> > mandarin
> > on patchwork. It's so fresh and exciting :-).
> >
>
> 谢谢。 I'm learning mainland Chinese for a few month now, I also learned that
> you
> use different symbols in Taiwan, which I don't know. 对不起。
Ha. Both symbols are welcome to me. :-)
>
> > >
> > > On 01/02/2023 13:28, Roger Lu (陸瑞傑) wrote:
> > > > Hi Matthias Sir,
> > > >
> > > > On Tue, 2023-01-31 at 14:19 +0100, Matthias Brugger wrote:
> > > > >
> > > > > On 11/01/2023 08:45, Roger Lu wrote:
> > > > > > In MediaTek HW design, svs and thermal both use the same clk source.
> > > > > > It means that svs clk reference count from CCF includes thermal
> > > > > > control
> > > > > > counts. That makes svs driver confuse whether it disabled svs's main
> > > > > > clk
> > > > > > or not from CCF's perspective and lead to turn off their shared clk
> > > > > > unexpectedly. Therefore, we add svs clk control APIs to make sure
> > > > > > svs's
> > > > > > main clk is controlled well by svs driver itself.
> > > > > >
> > > > > > Here is a NG example. Rely on CCF's reference count and cause
> > > > > > problem.
> > > > > >
> > > > > > thermal probe (clk ref = 1)
> > > > > > -> svs probe (clk ref = 2)
> > > > > > -> svs suspend (clk ref = 1)
> > > > > > -> thermal suspend (clk ref = 0)
> > > > > > -> thermal resume (clk ref = 1)
> > > > > > -> svs resume (encounter error, clk ref = 1)
> > > > > > -> svs suspend (clk ref = 0)
> > > > > > -> thermal suspend (Fail here, thermal HW control w/o clk)
> > > > > >
> > > > > > Fixes: a825d72f74a3 ("soc: mediatek: fix missing
> > > > > > clk_disable_unprepare()
> > > > > > on
> > > > > > err in svs_resume()")
> > > > > > Signed-off-by: Roger Lu <[email protected]>
> > > > >
> > > > > That looks wrong. Although I don't out of my mind, there should be a
> > > > > way
> > > > > to
> > > > > tell
> > > > > the clock framework that this clock is shared between several devices.
> > > > >
> > > > > I wonder if using clk_enable and clk_disable in svs_resume/suspend
> > > > > wouldn't
> > > > > be
> > > > > enough.
> > > >
> > > > Oh yes, Common Clock Framework (CCF) knows the clock shared between
> > > > several
> > > > devices and maintains clock "on/off" by reference count.
> > > >
> > >
> > > The thing is if you use clk_prepare_enable then the clock framework
> > > check's
> > > if
> > > the clock is already prepared, which could happen like you described in
> > > the
> > > svs_resume (encount error) case in the commit message. The question is,
> > > can't
> > > we
> > > just use clk_enable and clk_disable in resume/suspend and only prepare the
> > > clock
> > > in the probe function?
> >
> > We'll think if this can fix the problem. Thanks for the advice very much.
> >
> > >
> > > > We concern how to stop running svs_suspend() when svs clk is already
> > > > disabled by
> > > > svs_resume(). Take an example as below, if we refers to
> > > > __clk_is_enabled()
> > > > result for knowing svs clk status, it will return "true" all the time
> > > > because
> > > > thermal clk is still on. This causes the problem mentioned in commit
> > > > message.
> > > >
> > >
> > > I would expect that the kernel takes care that we can't enter a resume
> > > path
> > > for
> > > a device before the suspend path has finished. Honestly I don't really
> > > understand the problem here. It seems something different then what you
> > > described in the commit message.
> > >
> > > Please help me understand better.
> >
> > I see. This patch title needs to be changed to "avoid turning off svs clk
> > twice
> > unexpectedly" for pointing out the problem precisely. We saw a loophole that
> > svs
> > clk might be turned off in svs_resume() first and in svs_suspend() again
> > without
> > enabling svs clk during these the process. Therefore, we try to fix it by
> > this
> > patch. Below is our thinking process to explain how we got here.
> >
> > 1. (abandoned) We add __clk_is_enabled() check in svs_suspend() to prevent
> > svs
> > clk from being turned off twice when svs_resume() turned off svs clk in the
> > error-handling process. Nonetheless, we met the NG case in the commit
> > message.
> > 2. (current patch) We add svs clk control hint to understand if we need to
> > run
> > svs_suspend() or not if svs_resume() turned off svs clk before.
> >
>
> Did you had a look on the dev_pm_ops？ Maybe we can use suspend_late,
> resume_early to make sure there is no race condition. I wonder also if we
> can't
> make sure that this does not happen using device links. Sorry, I can't give
> better guidance on how to use this technologies, but I have the feeling we
> can
> fix this with existing infrastructure.
No, we didn't look on dev_pm_ops and it seems like another way to fix this issue
with __clk_is_enabled() check in svs_suspend(). Thanks for the advice again.
we'll keep looking for any possible answers to this issue.
Sincerely,
Roger Lu.