Re: [PATCH] MMC:dw_mmc:when enable biu_clk should check whether this clk is normal
From: Shawn Lin
Date: Sun Jan 04 2026 - 20:46:36 EST
在 2026/01/04 星期日 22:48, Sebastian Reichel 写道:
Hi,
On Sun, Jan 04, 2026 at 03:38:42PM +0800, Shawn Lin wrote:
Hi Sevastian,
在 2026/01/04 星期日 14:01, Sebastian Reichel 写道:
Hi,
On Sun, Jan 04, 2026 at 10:55:48AM +0800, Shawn Lin wrote:
在 2026/01/04 星期日 10:38, ping.gao 写道:
some vendor doesn't use biu_clk , enable it will panic
log is below:
I still don't understand the real issue you tried to solve.
host->biu_clk is got via devm_clk_get() in dw_mci_probe(). If
succeed, immediately it's enabled a few lines below. Otherwise,
dw_mci_probe() returns with failure. Seems you are not enabling it in
runtime callbacks, but why clk_prepare_enable(host->biu_clk) is able
to work when driver probed?
Based on the error and this patch, I believe what happens is:
1. dw_mci_probe() calls devm_clk_get(host->dev, "biu")
2. The clock does not exist on the platform from Ping Gao, so
host->biu_clk gets an error pointer assigned
3. "biu clock not available" gets printed at debug level
4. the error is ignored, but the error pointer is still assigned to
host->biu_clk
5. in the runtime PM handler clk_prepare_enable is called for
host->biu_clk and thus for an error pointer
Thanks for the details. For example,
1. dw_mci_exynos_probe() enables runtime PM before calling
dw_mci_pltfm_register()->dw_mci_probe().
2. biu_clk gets an error and propagate it to dw_mci_pltfm_register(),
The error propagation does not happen. This is the current code in
dw_mci_probe():
host->biu_clk = devm_clk_get(host->dev, "biu");
if (IS_ERR(host->biu_clk)) {
dev_dbg(host->dev, "biu clock not available\n");
ret = PTR_ERR(host->biu_clk);
if (ret == -EPROBE_DEFER)
return ret;
/* --- no return here --- */
} else {
ret = clk_prepare_enable(host->biu_clk);
if (ret) {
dev_err(host->dev, "failed to enable biu clock\n");
return ret;
}
}
This code continues probing for any error code that is not
-EPROBE_DEFER. My understanding is, that this has been done
intensionally, since the clock does not exist on some platforms.
I never thought of biu_clk isn't been provided as it's for accessing
register/fifo/interrupt/DMA requested by dwc databook. I was
suprised it's the optional case for some platforms. What I was thinking
is all based on that defer case.
Effectively this is an open coded devm_clk_get_optional with
the difference that you may not use any clk_* operations on the
error code. A minimal fix would be to assign "host->biu_clk = NULL;"
at the end of the error handling code, which would open code
more of the devm_clk_get_optional() behaviour.
Yes, if it's optional case for some platforms, we should use
devm_clk_get_optional() instead, both for biu_clk and ciu_clk.
Although I will also be surprised to see ciu_clk is also
optional, the code should be consistent.
Obviously there is no point in having this open coded and the
helper existed since ages, so I strongly suggest to just use
that. Either devm_clk_get_optional() or even better
devm_clk_get_optional_enabled(). This results in:
- less code, which is easier to understand
- fixing the bug at hand
- proper error handling for errors that are neither -EPROBE_DEFER
nor -ENOENT.
- better error/debug messaging behaviour for -EPROBE_DEFER when
using dev_err_probe() instead of the dev_dbg print.
Greetings,
-- Sebastian
3. runtime PM is disabled in dw_mci_exynos_probe() later if seeing error.
Does that mean runtime PM handler is called between step 2 and step 3?
If that is the case, how could we allow this happen because the the
controller's configuaration isn't been finished yet. Or maybe I
misunderstand your point?
Instead of hacking this up even further (as this patch does),
the proper fix is to cleanup the probe function. I believe
the complete biu clock handling in it could be reduced to this:
host->biu_clk = devm_clk_get_optional_enabled(host->dev, "biu");
ret = PTR_ERR_OR_ZERO(host->biu_clk);
if (ret)
return dev_err_probe(host->dev, ret, "failed to get biu clock\n");
That will assign NULL to host->biu_clk when there is no clock
defined and it is fine to run clk_prepare_enable() with NULL
as clock pointer. Note, that this will handle the clock enabling
as part of devm, so you will also have to drop the
clk_disable_unprepare() calls from the err_clk_biu goto label
and the extra call from dw_mci_remove().
Something similar should probably be done for host->ciu_clk, which
seems to have the same incorrect logic.
Greetings,
-- Sebastian
[ 438.400868] [7: binder:436_2: 4998] Unable to handle kernel paging request at virtual address fffffffffffffffe
[ 438.400877] [7: binder:436_2: 4998] Mem abort info:
[ 438.400881] [7: binder:436_2: 4998] ESR = 0x0000000096000005
[ 438.400887] [7: binder:436_2: 4998] EC = 0x25: DABT (current EL), IL = 32 bits
[ 438.400894] [7: binder:436_2: 4998] SET = 0, FnV = 0
[ 438.400899] [7: binder:436_2: 4998] EA = 0, S1PTW = 0
[ 438.400904] [7: binder:436_2: 4998] FSC = 0x05: level 1 translation fault
...
[ 438.409424] [7: binder:436_2: 4998] Call trace:
[ 438.409429] [7: binder:436_2: 4998] clk_prepare+0x10/0x24
[ 438.409439] [7: binder:436_2: 4998] dw_mci_runtime_resume+0x50/0x2d8 [dw_mmc_samsung cd210e210975263404c28fc89778f369f8398f0c]
[ 438.409471] [7: binder:436_2: 4998] dw_mci_exynos_runtime_resume+0x18/0x58 [dw_mmc_exynos_samsung 2735a594c7c9c9e8c65b0b87523fbf70dcaabfff]
[ 438.409496] [7: binder:436_2: 4998] pm_generic_runtime_resume+0x40/0x58
[ 438.409506] [7: binder:436_2: 4998] pm_runtime_force_resume+0x9c/0x134
[ 438.409517] [7: binder:436_2: 4998] platform_pm_resume+0x40/0x8c
[ 438.409529] [7: binder:436_2: 4998] dpm_run_callback+0x64/0x230
[ 438.409540] [7: binder:436_2: 4998] __device_resume+0x1d8/0x394
[ 438.409551] [7: binder:436_2: 4998] dpm_resume+0x110/0x2b8
[ 438.409561] [7: binder:436_2: 4998] dpm_resume_end+0x1c/0x38
[ 438.409570] [7: binder:436_2: 4998] suspend_devices_and_enter+0x828/0xab0
[ 438.409582] [7: binder:436_2: 4998] pm_suspend+0x334/0x618
[ 438.409592] [7: binder:436_2: 4998] state_store+0x104/0x144
[ 438.409601] [7: binder:436_2: 4998] kobj_attr_store+0x30/0x48
[ 438.409610] [7: binder:436_2: 4998] sysfs_kf_write+0x54/0x6c
[ 438.409619] [7: binder:436_2: 4998] kernfs_fop_write_iter+0x104/0x1a8
[ 438.409628] [7: binder:436_2: 4998] vfs_write+0x24c/0x2f4
[ 438.409640] [7: binder:436_2: 4998] ksys_write+0x78/0xe8
[ 438.409652] [7: binder:436_2: 4998] __arm64_sys_write+0x1c/0x2c
[ 438.409664] [7: binder:436_2: 4998] invoke_syscall+0x58/0x114
[ 438.409676] [7: binder:436_2: 4998] el0_svc_common+0xac/0xe0
[ 438.409687] [7: binder:436_2: 4998] do_el0_svc+0x1c/0x28
[ 438.409698] [7: binder:436_2: 4998] el0_svc+0x38/0x68
[ 438.409705] [7: binder:436_2: 4998] el0t_64_sync_handler+0x68/0xbc
[ 438.409712] [7: binder:436_2: 4998] el0t_64_sync+0x1a8/0x1ac
Signed-off-by: ping.gao <ping.gao@xxxxxxxxxxx>
---
drivers/mmc/host/dw_mmc.c | 12 +++++++++---
1 file changed, 9 insertions(+), 3 deletions(-)
diff --git a/drivers/mmc/host/dw_mmc.c b/drivers/mmc/host/dw_mmc.c
index c5db92bbb094..61f6986f15ef 100644
--- a/drivers/mmc/host/dw_mmc.c
+++ b/drivers/mmc/host/dw_mmc.c
@@ -3633,9 +3633,15 @@ int dw_mci_runtime_resume(struct device *dev)
if (host->slot &&
(mmc_host_can_gpio_cd(host->slot->mmc) ||
!mmc_card_is_removable(host->slot->mmc))) {
- ret = clk_prepare_enable(host->biu_clk);
- if (ret)
- return ret;
+ if (IS_ERR(host->biu_clk)) {
+ dev_dbg(host->dev, "biu clock not available\n");
+ } else {
+ ret = clk_prepare_enable(host->biu_clk);
+ if (ret) {
+ dev_err(host->dev, "failed to enable biu clock\n");
+ goto err;
+ }
+ }
}
ret = clk_prepare_enable(host->ciu_clk);