pmdomain: mediatek: fix race conditions with genpd
Subsystems (2)
PATCH 1/1 pmdomain: mediatek: fix race conditions with genpd
If the power domains are registered first with genpd and *after that* the driver attempts to power them on in the probe sequence, then it is possible that a race condition occurs if genpd tries to power them on in the same time.
The same is valid for powering them off before unregistering them from genpd.
Attempt to fix race conditions by first removing the domains from genpd and *after that* powering down domains.
Also first power up the domains and *after that* register them to genpd.
---
This comes as another way to fix the problem as described in this thread:
https://lore.kernel.org/linux-arm-kernel/[email protected]/
I have not been able to reproduce the problem with either fix anymore (so far).
I have a few doubts about this one though, if I really covered the way it's supposed to work, and registering the pmdomains in the recursive function in the reversed order has any side effect or if it does not work correctly.
Tested on mt8186 where it appears to be fine.
Eugen
drivers/pmdomain/mediatek/mtk-pm-domains.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/pmdomain/mediatek/mtk-pm-domains.c +7 -8
@@ -561,6 +561,11 @@ static int scpsys_add_subdomain(struct scpsys *scpsys, struct device_node *paren goto err_put_node; } +/* recursive call to add all subdomains */ + ret =scpsys_add_subdomain (scpsys, child); +if (ret) + goto err_put_node; + ret =pm_genpd_add_subdomain (parent_pd, child_pd);if (ret) { dev_err(scpsys->dev, "failed to add %s subdomain to parent %s\n",
@@ -570,11 +575,6 @@ static int scpsys_add_subdomain(struct scpsys *scpsys, struct device_node *parendev_dbg (scpsys->dev,"%s add subdomain: %s\n" , parent_pd->name, child_pd->name); } - -/* recursive call to add all subdomains */ - ret =scpsys_add_subdomain (scpsys, child); -if (ret) - goto err_put_node; }return 0 ;
@@ -588,9 +588,6 @@ static void scpsys_remove_one_domain(struct scpsys_domain *pd) {int ret; -if (scpsys_domain_is_on (pd)) -scpsys_power_off (&pd->genpd); -/* * We're in the error cleanup already, so we only complain, * but won't emit another error on top of the original one.
@@ -600,6 +597,8 @@ static void scpsys_remove_one_domain(struct scpsys_domain *pd)dev_err (pd->scpsys->dev,"failed to remove domain '%s' : %d - state may be inconsistent\n" , pd->genpd.name, ret); +if (scpsys_domain_is_on (pd)) +scpsys_power_off (&pd->genpd);clk_bulk_put (pd->num_clks, pd->clks);clk_bulk_put (pd->num_subsys_clks, pd->subsys_clks);
Signature · 1 line
2.34.1
Quoted 7 lines · Eugen Hristev
If the power domains are registered first with genpd and *after that* the driver attempts to power them on in the probe sequence, then it is possible that a race condition occurs if genpd tries to power them on in the same time.
The same is valid for powering them off before unregistering them from genpd.
Right. When the PM domain has been registered with genpd, attempts to power-on/off the PM domain need to be synchronized with genpd.
Quoted 4 lines
Attempt to fix race conditions by first removing the domains from genpd and *after that* powering down domains.
Also first power up the domains and *after that* register them to genpd.
This seems like a reasonable approach to me.
Quoted 3 lines
Fixes 59b644b01cf4 ("soc: mediatek: Add MediaTek SCPSYS power domains")Signed-off-by Eugen Hristev <[email protected]>
Applied for fixes and by adding a stable tag, thanks! Although, please, see some more comments below.
Quoted 5 lines
---
This comes as another way to fix the problem as described in this thread:
https://lore.kernel.org/linux-arm-kernel/[email protected]/
Quoted 8 lines
I have not been able to reproduce the problem with either fix anymore (so far).
I have a few doubts about this one though, if I really covered the way it's supposed to work, and registering the pmdomains in the recursive function in the reversed order has any side effect or if it does not work correctly.
Tested on mt8186 where it appears to be fine.
I had a quick look at the code in the driver and a few things caught my eyes.
*) The error path in scpsys_probe() doesn't seem to handle removal of the link between parent/child-domains (subdomains).
**) An option that might simplify the code and error path too, could be to convert into using of_genpd_add|remove_subdomain() in favor or pm_genpd_add|remove_subdomain().
Kind regards
Uffe
Quoted 57 lines
Eugen
drivers/pmdomain/mediatek/mtk-pm-domains.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)drivers/pmdomain/mediatek/mtk-pm-domains.c +7
-9@@ -561,6 +561,11 @@ static int scpsys_add_subdomain(struct scpsys *scpsys, struct device_node *paren goto err_put_node; } +/* recursive call to add all subdomains */ + ret =scpsys_add_subdomain (scpsys, child); +if (ret) + goto err_put_node; + ret =pm_genpd_add_subdomain (parent_pd, child_pd);if (ret) { dev_err(scpsys->dev, "failed to add %s subdomain to parent %s\n",@@ -570,11 +575,6 @@ static int scpsys_add_subdomain(struct scpsys *scpsys, struct device_node *parendev_dbg (scpsys->dev,"%s add subdomain: %s\n" , parent_pd->name, child_pd->name); } - -/* recursive call to add all subdomains */ - ret =scpsys_add_subdomain (scpsys, child); -if (ret) - goto err_put_node; }return 0 ;@@ -588,9 +588,6 @@ static void scpsys_remove_one_domain(struct scpsys_domain *pd) {int ret; -if (scpsys_domain_is_on (pd)) -scpsys_power_off (&pd->genpd); -/* * We're in the error cleanup already, so we only complain, * but won't emit another error on top of the original one. @@ -600,6 +597,8 @@ static void scpsys_remove_one_domain(struct scpsys_domain *pd)dev_err (pd->scpsys->dev,"failed to remove domain '%s' : %d - state may be inconsistent\n" , pd->genpd.name, ret); +if (scpsys_domain_is_on (pd)) +scpsys_power_off (&pd->genpd);clk_bulk_put (pd->num_clks, pd->clks);clk_bulk_put (pd->num_subsys_clks, pd->subsys_clks); --2.34.1