Improvements to the recent clock rework by marex.
Subsystems (2)
cover Improvements to the recent clock rework by marex.
Symptoms:
I ran into an issue with the generic ccf_clk_divider clock driver. ccf_clk_divider assumes that the pointer passed to clk_divider_recalc_rate() (https://github.com/u-boot/u-boot/blob/master/drivers/clk/clk-divider.c#L80) is the same that was kzalloc()'d during clock registration. This assumption was breaking when calling clk_get_rate() on a clock retrieved via clk_get_by_index(), since clk_get_by_index() returns a copy of struct clk rather than the original pointer. For example, in https://github.com/u-boot/u-boot/blob/master/drivers/serial/serial_s5p.c#L131 the serial driver retrieves a clock using clk_get_by_index(), and then indirectly passes a pointer to stack-allocated struct clk to clk_divider_recalc_rate(), which breaks the ccf_clk_divider driver.
Description:
The Clock API relies on the fact that you call it's functions with a pointer to the kzalloc()'d struct clk, and not a copy of it. To ensure this, existing clock drivers use ccf_clk_get_rate() and clk_get_by_id().
The clock rework path missed this fact, and instead was returning names and pointers for the root device.
Testing:
I tested this on Samsung S10 (exynos9820 SoC) with a custom clock driver and a device tree. Unfortunately, I don't own any other exynos devices supported by u-boot.
chiffathefox (2):
clk: clk_resolve_parent_clk(): resolve external clocks by id
clk: exynos: do not reparent requested clock
drivers/clk/clk-uclass.c | 7 ++++++- drivers/clk/exynos/clk.c | 1 - 2 files changed, 6 insertions(+), 2 deletions(-)
Signature · 4 lines
2.34.1 base-commit: d503633a36767d756c7de28305cf0de79440cbc0 branch: clk-rework-improvements
PATCH 1/2 clk: clk_resolve_parent_clk(): resolve external clocks by id
When referencing clocks in DTS registered by another driver, clk_resolve_parent_clk() is supposed to return the name of the child device that provides the requested clock. However, in reality, it returns the name of the parent DTB node (e.g. clock-controller@1a240000). Such devices, for instance the exynos7870-cmu-peri driver, do not store any clock information in their dev->uclass_priv_ because they expect you to call dev->ops->get_rate with the appropriate clk->id. So when you try to resolve a clock that depends on an externally referenced clock, the resolution fails on the parent clock because it doesn't have a dev->uclass_priv_. With this change clk_resolve_parent_clk() will take an extra step to actually resolve the child device.
---
drivers/clk/clk-uclass.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/clk/clk-uclass.c +6 -1
@@ -430,6 +430,7 @@ clk_resolve_parent_clk(struct udevice *dev, const char *name) {struct udevice *parent;struct clk clk; +struct clk *c;int ret; ret =uclass_get_device_by_name (UCLASS_CLK , name, &parent);
@@ -440,7 +441,11 @@ clk_resolve_parent_clk(struct udevice *dev, const char *name)if (!clk.dev)return name; -return clk.dev->name; + ret =clk_get_by_id (clk.id, &c); +if (ret) +return name; + +return c->dev->name; }int clk_release_all (struct clk *clk,unsigned int count)
Signature · 1 line
2.34.1
Re: [PATCH 1/2] clk: clk_resolve_parent_clk(): resolve external clocks by id
Quoted 35 lines · chiffathefox
When referencing clocks in DTS registered by another driver,
clk_resolve_parent_clk()is supposed to return the name of the child device that provides the requested clock. However, in reality, it returns the name of the parent DTB node (e.g. clock-controller@1a240000). Such devices, for instance the exynos7870-cmu-peri driver, do not store any clock information in theirdev->uclass_priv_because they expect you to calldev->ops->get_ratewith the appropriateclk->id. So when you try to resolve a clock that depends on an externally referenced clock, the resolution fails on the parent clock because it doesn't have adev->uclass_priv_. With this changeclk_resolve_parent_clk()will take an extra step to actually resolve the child device.Signed-off-by chiffathefox <[email protected]>---
drivers/clk/clk-uclass.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)drivers/clk/clk-uclass.c +1
-0@@ -430,6 +430,7 @@ clk_resolve_parent_clk(struct udevice *dev, const char *name) {struct udevice *parent;struct clk clk; +struct clk *c;int ret; ret =uclass_get_device_by_name (UCLASS_CLK , name, &parent);@@ -440,7 +441,11 @@ clk_resolve_parent_clk(struct udevice *dev, const char *name)if (!clk.dev)return name;
drivers/clk/clk-uclass.c clk_resolve_parent_clk(struct udevice *dev, const char *name) · line 451
Quoted 10 lines
drivers/clk/clk-uclass.c +5
-1-return clk.dev->name; + ret =clk_get_by_id (clk.id, &c); +if (ret) +return name; + +return c->dev->name; }int clk_release_all (struct clk *clk,unsigned int count)
The issue description looks very similar to what this commit fixes:
0e9055b1488f ("clk: scmi: Fix priv initialization in scmi_clk_gate()")
Is this related ?
Can you share an example / test which triggers the problem ? (ideally, use sandbox or sandbox64, that would make it reproducible for others too)
Thank you
+CC Peng/Patrice.
Re: [PATCH 1/2] clk: clk_resolve_parent_clk(): resolve external clocks by id
In reply to chiffathefoxQuoted 14 lines · chiffathefox
When referencing clocks in DTS registered by another driver,
clk_resolve_parent_clk()is supposed to return the name of the child device that provides the requested clock. However, in reality, it returns the name of the parent DTB node (e.g. clock-controller@1a240000). Such devices, for instance the exynos7870-cmu-peri driver, do not store any clock information in theirdev->uclass_priv_because they expect you to calldev->ops->get_ratewith the appropriateclk->id. So when you try to resolve a clock that depends on an externally referenced clock, the resolution fails on the parent clock because it doesn't have adev->uclass_priv_. With this changeclk_resolve_parent_clk()will take an extra step to actually resolve the child device.
Quoted 1 line
Signed-off-by chiffathefox <[email protected]>
PATCH 2/2 clk: exynos: do not reparent requested clock
In reply to chiffathefoxIt would appear the original developer of this driver reparented child clock devices to fix the issue described in clk: clk_resolve_parent_clk(): resolve external clocks by id. This hacky fix was the cause of invalid memory access when calling clk_get_rate() on a clock retrieved via clk_get_by_index(). clk_get_by_index() returns a copy of the struct clk and when passed to clk_get_rate() it may be forwarded to clk_divider_recalc_rate(), which will cause the latter to run container_of() on a pointer from caller's stack instead of the pointer that was kzalloc()'d by the clk-divider driver.
---
drivers/clk/exynos/clk.c | 1 - 1 file changed, 1 deletion(-)
drivers/clk/exynos/clk.c +0 -1
@@ -19,7 +19,6 @@ int samsung_clk_request(struct clk *clk)if (ret)return ret; - clk->dev = c->dev;return 0 ; }
Signature · 1 line
2.34.1
Re: [PATCH 2/2] clk: exynos: do not reparent requested clock
Quoted 11 lines · chiffathefox
It would appear the original developer of this driver reparented child clock devices to fix the issue described in
clk: clk_resolve_parent_clk(): resolve external clocks by id. This hacky fix was the cause of invalid memory access when callingclk_get_rate()on a clock retrieved viaclk_get_by_index().clk_get_by_index()returns a copy of thestruct clkand when passed toclk_get_rate()it may be forwarded toclk_divider_recalc_rate(), which will cause the latter to runcontainer_of()on a pointer from caller's stack instead of the pointer that was kzalloc()'d by the clk-divider driver.
Quoted 1 line
Signed-off-by chiffathefox <[email protected]>