Current implementation of scpi_clk_add just print a warning when clock
fails to register but then keep going as if nothing happened. The
provider is then registered with bogus data.
This may latter lead to an Oops in __clk_create_clk when
hlist_add_head(&clk->clks_node, &hw->core->clks) is called.
This patch fixes the issue and errors if a clock fails to register.
Fixes: cd52c2a4b5c4 ("clk: add support for clocks provided by SCP(System Control Processor)")
Signed-off-by: Jerome Brunet <jbrunet@baylibre.com>
---
drivers/clk/clk-scpi.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
Current implementation of scpi_clk_add just print a warning when clock
fails to register but then keep going as if nothing happened. The
provider is then registered with bogus data.
This may latter lead to an Oops in __clk_create_clk when
hlist_add_head(&clk->clks_node, &hw->core->clks) is called.
What's the path of this call ? I see one in devm_clk_hw_register
but that's one which failed.
Also one of the reason for keeping it continuing is, if firmware fails
on some non-critical clock, that's fine rather than punishing the entire
set of clocks and may even fail the boot.
--
Regards,
Sudeep
On Wed, 2017-06-28 at 16:04 +0100, Sudeep Holla wrote:
On 28/06/17 14:53, Jerome Brunet wrote:
quoted
Current implementation of scpi_clk_add just print a warning when clock
fails to register but then keep going as if nothing happened. The
provider is then registered with bogus data.
This may latter lead to an Oops in __clk_create_clk when
hlist_add_head(&clk->clks_node, &hw->core->clks) is called.
What's the path of this call ? I see one in devm_clk_hw_register
but that's one which failed.
bL cpu freq driver requesting the cpu clock, which failed to register. Here the
Oops call trace:
[????2.202284] [<ffff00000849a058>] __clk_create_clk.part.18+0x68/0xb0
[????2.208494] [<ffff00000849ac2c>] __of_clk_get_from_provider+0xfc/0x140
[????2.214962] [<ffff000008496c28>] __of_clk_get_by_name+0x100/0x118
[????2.220999] [<ffff000008496c94>] clk_get+0x2c/0x78
[????2.225744] [<ffff000008570110>] dev_pm_opp_get_opp_table+0xb0/0x118
[????2.232039] [<ffff000008570940>] dev_pm_opp_add+0x20/0x68
[????2.237388] [<ffff0000087a0f30>] scpi_init_opp_table+0xa8/0x188
[????2.243252] [<ffff0000087a0558>] _get_cluster_clk_and_freq_table+0x80/0x180
[????2.250151] [<ffff0000087a0a48>] bL_cpufreq_init+0x3f0/0x480
[????2.255758] [<ffff00000879eed8>] cpufreq_online+0xc0/0x658
[????2.261191] [<ffff00000879f500>] cpufreq_add_dev+0x78/0x88
[????2.266625] [<ffff00000855c2c4>] subsys_interface_register+0x84/0xc8
[????2.272922] [<ffff00000879e330>] cpufreq_register_driver+0x138/0x1b8
[????2.279218] [<ffff0000087a0b4c>] bL_cpufreq_register+0x74/0x120
[????2.285083] [<ffff0000087a1038>] scpi_cpufreq_probe+0x28/0x38
[????2.290776] [<ffff00000855fbf0>] platform_drv_probe+0x50/0xb8
[????2.296468] [<ffff00000855dd84>] driver_probe_device+0x21c/0x2d8
But that's not the point. The point is there is path in clk-scpi driver which
registers uninitialized data in the clock provider. That's not good.?
Also one of the reason for keeping it continuing is, if firmware fails
on some non-critical clock, that's fine rather than punishing the entire
set of clocks and may even fail the boot.
I understand, but you have no way to know whether a clock is critical or not so
this explanation looks a bit weak, plus it does not keep the boot from failing
... not for me at least.
As explained this approach is registering corrupt data in the provider when
failing. It makes the kernel Oops, it shall be fixed.
If you have a better solution later on, I don't think there would be any problem
to revert this patch.
On Wed, 2017-06-28 at 16:04 +0100, Sudeep Holla wrote:
quoted
On 28/06/17 14:53, Jerome Brunet wrote:
quoted
Current implementation of scpi_clk_add just print a warning when clock
fails to register but then keep going as if nothing happened. The
provider is then registered with bogus data.
This may latter lead to an Oops in __clk_create_clk when
hlist_add_head(&clk->clks_node, &hw->core->clks) is called.
What's the path of this call ? I see one in devm_clk_hw_register
but that's one which failed.
Thanks for this stack. I just worked out the same path now. I did come
up with the patch as below. That should work if my understanding is correct.
But that's not the point. The point is there is path in clk-scpi driver which
registers uninitialized data in the clock provider. That's not good.
quoted
Also one of the reason for keeping it continuing is, if firmware fails
on some non-critical clock, that's fine rather than punishing the entire
set of clocks and may even fail the boot.
I understand, but you have no way to know whether a clock is critical or not so
this explanation looks a bit weak, plus it does not keep the boot from failing
... not for me at least.
As explained this approach is registering corrupt data in the provider when
failing. It makes the kernel Oops, it shall be fixed.
Agreed, I want to look at ways to fix that, hence requested you more data.
If you have a better solution later on, I don't think there would be any problem
to revert this patch.
Sure I am not against the patch as a fix. I was just trying to better
understand the problem. I had seen the usefulness of skipping on Juno
platforms
in earlier days. Also I am now working on the new SCMI[1] specification
and just want to cover this.
---
On Wed, 2017-06-28 at 16:52 +0100, Sudeep Holla wrote:
On 28/06/17 16:38, Jerome Brunet wrote:
quoted
On Wed, 2017-06-28 at 16:04 +0100, Sudeep Holla wrote:
quoted
On 28/06/17 14:53, Jerome Brunet wrote:
quoted
Current implementation of scpi_clk_add just print a warning when clock
fails to register but then keep going as if nothing happened. The
provider is then registered with bogus data.
This may latter lead to an Oops in __clk_create_clk when
hlist_add_head(&clk->clks_node, &hw->core->clks) is called.
What's the path of this call ? I see one in devm_clk_hw_register
but that's one which failed.
bL cpu freq driver requesting the cpu clock, which failed to register. Here
the
Oops call trace:
[????2.202284] [<ffff00000849a058>] __clk_create_clk.part.18+0x68/0xb0
[????2.208494] [<ffff00000849ac2c>] __of_clk_get_from_provider+0xfc/0x140
[????2.214962] [<ffff000008496c28>] __of_clk_get_by_name+0x100/0x118
[????2.220999] [<ffff000008496c94>] clk_get+0x2c/0x78
[????2.225744] [<ffff000008570110>] dev_pm_opp_get_opp_table+0xb0/0x118
[????2.232039] [<ffff000008570940>] dev_pm_opp_add+0x20/0x68
[????2.237388] [<ffff0000087a0f30>] scpi_init_opp_table+0xa8/0x188
[????2.243252] [<ffff0000087a0558>]
_get_cluster_clk_and_freq_table+0x80/0x180
[????2.250151] [<ffff0000087a0a48>] bL_cpufreq_init+0x3f0/0x480
[????2.255758] [<ffff00000879eed8>] cpufreq_online+0xc0/0x658
[????2.261191] [<ffff00000879f500>] cpufreq_add_dev+0x78/0x88
[????2.266625] [<ffff00000855c2c4>] subsys_interface_register+0x84/0xc8
[????2.272922] [<ffff00000879e330>] cpufreq_register_driver+0x138/0x1b8
[????2.279218] [<ffff0000087a0b4c>] bL_cpufreq_register+0x74/0x120
[????2.285083] [<ffff0000087a1038>] scpi_cpufreq_probe+0x28/0x38
[????2.290776] [<ffff00000855fbf0>] platform_drv_probe+0x50/0xb8
[????2.296468] [<ffff00000855dd84>] driver_probe_device+0x21c/0x2d8
Thanks for this stack. I just worked out the same path now. I did come
up with the patch as below. That should work if my understanding is correct.
I tried.
It does not work unfortunately. Still crashes but somewhere else:
[????2.301482] [<ffff00000849e67c>] scpi_of_clk_src_get+0x14/0x58
[????2.307261] [<ffff000008495f40>] __of_clk_get_by_name+0x100/0x118
[????2.313297] [<ffff000008495fac>] clk_get+0x2c/0x78
[????2.318044] [<ffff00000856f4d0>] dev_pm_opp_get_opp_table+0xb0/0x118
[????2.324338] [<ffff00000856fd00>] dev_pm_opp_add+0x20/0x68
[????2.329687] [<ffff0000087a04f8>] scpi_init_opp_table+0xa8/0x188
[????2.335550] [<ffff00000879fb20>] _get_cluster_clk_and_freq_table+0x80/0x180
[????2.342450] [<ffff0000087a0010>] bL_cpufreq_init+0x3f0/0x480
[????2.348056] [<ffff00000879e4a0>] cpufreq_online+0xc0/0x658
[????2.353490] [<ffff00000879eac8>] cpufreq_add_dev+0x78/0x88
[????2.358924] [<ffff00000855b684>] subsys_interface_register+0x84/0xc8
[????2.365220] [<ffff00000879d8f8>] cpufreq_register_driver+0x138/0x1b8
[????2.371516] [<ffff0000087a0114>] bL_cpufreq_register+0x74/0x120
[????2.377381] [<ffff0000087a0600>] scpi_cpufreq_probe+0x28/0x38
[????2.383076] [<ffff00000855efb0>] platform_drv_probe+0x50/0xb8
[????2.388766] [<ffff00000855d144>] driver_probe_device+0x21c/0x2d8
I have not looked at ALL the clock providers, but I have seen a few and I don't
remember seeing any which fails, at some point, to register a clocks and still
register successfully.
It seems strange to continue with a broken controller.
quoted hunk
quoted
But that's not the point. The point is there is path in clk-scpi driver
which
registers uninitialized data in the clock provider. That's not good.?
quoted
Also one of the reason for keeping it continuing is, if firmware fails
on some non-critical clock, that's fine rather than punishing the entire
set of clocks and may even fail the boot.
I understand, but you have no way to know whether a clock is critical or not
so?
this explanation looks a bit weak, plus it does not keep the boot from
failing
... not for me at least.
As explained this approach is registering corrupt data in the provider when
failing. It makes the kernel Oops, it shall be fixed.
Agreed, I want to look at ways to fix that, hence requested you more data.
quoted
If you have a better solution later on, I don't think there would be any
problem
to revert this patch.
Sure I am not against the patch as a fix. I was just trying to better
understand the problem. I had seen the usefulness of skipping on Juno
platforms
in earlier days. Also I am now working on the new SCMI[1] specification
and just want to cover this.
---
Looks like a different route and I know why. I have added an extra check
now which should work if I have not missed anything more.
I have not looked at ALL the clock providers, but I have seen a few and I don't
remember seeing any which fails, at some point, to register a clocks and still
register successfully.
No problem, as I said I am fine with the patch you sent as a fix for now
but just curious to know what are the issues to be fixed to continue
supporting that feature. Please bear with me.
It seems strange to continue with a broken controller.
I would have agreed if it was single driver or h/w controlled by Linux.
Since it's in the firmware, we should allow the working clocks/opps to
work though few are broken. It's not good if we had to disable
everything if some piece of firmware is not yet ready or broken.
But again, we can get it working later, for now, I am fine with you patch.
Regards,
Sudeep
---
From: Stephen Boyd <hidden> Date: 2017-06-28 22:33:43
On 06/28, Sudeep Holla wrote:
everything if some piece of firmware is not yet ready or broken.
But again, we can get it working later, for now, I am fine with you patch.
So that's an acked-by, reviewed-by tag for the original patch?
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
On Wed, 2017-06-28 at 18:07 +0100, Sudeep Holla wrote:
On 28/06/17 17:46, Jerome Brunet wrote:
quoted
On Wed, 2017-06-28 at 16:52 +0100, Sudeep Holla wrote:
[..]
quoted
quoted
Thanks for this stack. I just worked out the same path now. I did come
up with the patch as below. That should work if my understanding is
correct.
I tried.
Thanks.
quoted
It does not work unfortunately. Still crashes but somewhere else:
[????2.301482] [<ffff00000849e67c>] scpi_of_clk_src_get+0x14/0x58
[????2.307261] [<ffff000008495f40>] __of_clk_get_by_name+0x100/0x118
[????2.313297] [<ffff000008495fac>] clk_get+0x2c/0x78
[????2.318044] [<ffff00000856f4d0>] dev_pm_opp_get_opp_table+0xb0/0x118
[????2.324338] [<ffff00000856fd00>] dev_pm_opp_add+0x20/0x68
[????2.329687] [<ffff0000087a04f8>] scpi_init_opp_table+0xa8/0x188
[????2.335550] [<ffff00000879fb20>]
_get_cluster_clk_and_freq_table+0x80/0x180
[????2.342450] [<ffff0000087a0010>] bL_cpufreq_init+0x3f0/0x480
[????2.348056] [<ffff00000879e4a0>] cpufreq_online+0xc0/0x658
[????2.353490] [<ffff00000879eac8>] cpufreq_add_dev+0x78/0x88
[????2.358924] [<ffff00000855b684>] subsys_interface_register+0x84/0xc8
[????2.365220] [<ffff00000879d8f8>] cpufreq_register_driver+0x138/0x1b8
[????2.371516] [<ffff0000087a0114>] bL_cpufreq_register+0x74/0x120
[????2.377381] [<ffff0000087a0600>] scpi_cpufreq_probe+0x28/0x38
[????2.383076] [<ffff00000855efb0>] platform_drv_probe+0x50/0xb8
[????2.388766] [<ffff00000855d144>] driver_probe_device+0x21c/0x2d8
Looks like a different route and I know why. I have added an extra check
now which should work if I have not missed anything more.
quoted
I have not looked at ALL the clock providers, but I have seen a few and I
don't
remember seeing any which fails, at some point, to register a clocks and
still
register successfully.?
No problem, as I said I am fine with the patch you sent as a fix for now
but just curious to know what are the issues to be fixed to continue
supporting that feature. Please bear with me.
I am :) and I understand what you are trying to do, having a degraded clock
provider is better than nothing according to you, correct?
I'm wondering whether this is correct or not, that why I'm challenging this a
bit.
If you failed to register an scpi clock it is probably because the communication
with the FW is not working, or at least 'not that good', right ?
If for some reason, you manage to register some other clocks from the same FW,
how confident can you be that communication will be ok for them ? that the
settings you request will be applied correctly ?
Is it possible that you may be causing more harm/damage playing with a broken HW
?
quoted
It seems strange to continue with a broken controller.
I would have agreed if it was single driver or h/w controlled by Linux.
Since it's in the firmware, we should allow the working clocks/opps to
work though few are broken. It's not good if we had to disable
everything if some piece of firmware is not yet ready or broken.
But again, we can get it working later, for now, I am fine with you patch.
I tried your last version, and it does not Oops, at least not for me.
The end result still looks odd to me:
[????1.115219] scpi_clocks scpi:clocks: failed to register clock 'vcpu'
[????1.159490] cpu cpu0: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 0, cluster: 0
[????1.162986] cpu cpu0: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[????1.170945] cpu cpu1: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 1, cluster: 0
[????1.179634] cpu cpu1: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[????1.187654] cpu cpu2: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 2, cluster: 0
[????1.196284] cpu cpu2: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[????1.204375] cpu cpu3: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 3, cluster: 0
[????1.212911] cpu cpu3: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[????1.220612] arm_big_little: bL_cpufreq_register: Registered platform driver:
scpi
So now, I have an scpi clock provider which registers successfully but fails to
register its only clock. As a consequence, I also have a cpufreq driver which
manages to register but has no clock cpu clock to drive ...
@@ -245,11 +245,14 @@ static int scpi_clk_add(struct device *dev, struct
device_node *np,
????????????????sclk->id = val;
????????????????err = scpi_clk_ops_init(dev, match, sclk, name);
-???????????????if (err)
+???????????????if (err) {
????????????????????????dev_err(dev, "failed to register clock '%s'\n",
name);
-???????????????else
+???????????????????????clk_data->clk[idx] = NULL;
+???????????????????????devm_kfree(dev, sclk);
+???????????????} else {
????????????????????????dev_dbg(dev, "Registered clock '%s'\n", name);
-???????????????clk_data->clk[idx] = sclk;
+???????????????????????clk_data->clk[idx] = sclk;
+???????????????}
????????}
????????return of_clk_add_hw_provider(np, scpi_of_clk_src_get, clk_data);
--
To unsubscribe from this list: send the line "unsubscribe linux-clk" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at??http://vger.kernel.org/majordomo-info.html
Hi Jerome,
Thanks for the fix.
On 28/06/17 14:53, Jerome Brunet wrote:
Current implementation of scpi_clk_add just print a warning when clock
fails to register but then keep going as if nothing happened. The
provider is then registered with bogus data.
This may latter lead to an Oops in __clk_create_clk when
hlist_add_head(&clk->clks_node, &hw->core->clks) is called.
This patch fixes the issue and errors if a clock fails to register.
Fixes: cd52c2a4b5c4 ("clk: add support for clocks provided by SCP(System Control Processor)")
Looks like a different route and I know why. I have added an extra check
now which should work if I have not missed anything more.
quoted
I have not looked at ALL the clock providers, but I have seen a few and I
don't
remember seeing any which fails, at some point, to register a clocks and
still
register successfully.
No problem, as I said I am fine with the patch you sent as a fix for now
but just curious to know what are the issues to be fixed to continue
supporting that feature. Please bear with me.
I am :) and I understand what you are trying to do, having a degraded clock
provider is better than nothing according to you, correct?
I'm wondering whether this is correct or not, that why I'm challenging this a
bit.
Fair enough. But the situation I had on my platform is that it provides
DVFS support for 2 CPU clusters and 1 GPU domain. I didn't want to block
using CPUFreq until GPU DVFS was properly supported in the firmware.
I had similar situation with the clock and hence I allowed it to continue.
If you failed to register an scpi clock it is probably because the communication
with the FW is not working, or at least 'not that good', right ?
Not exactly, what if the error is for that particular clock. That's my
point. If we have reached so far means the communication is fine. Just a
fault piece of hardware which may not be critical.
If for some reason, you manage to register some other clocks from the same FW,
how confident can you be that communication will be ok for them ? that the
settings you request will be applied correctly ?
Not sure, I am not registering the clock. Think SCPI as a single clock
provider with multiple clock outputs. You don't want to disable it
entirely if one of the clock outputs have problem. That's my counter
argument.
Is it possible that you may be causing more harm/damage playing with a broken HW
?
Not sure how if we are not registering that clock output from the h/w
clock provider perspective.
quoted
quoted
It seems strange to continue with a broken controller.
I would have agreed if it was single driver or h/w controlled by Linux.
Since it's in the firmware, we should allow the working clocks/opps to
work though few are broken. It's not good if we had to disable
everything if some piece of firmware is not yet ready or broken.
But again, we can get it working later, for now, I am fine with you patch.
I tried your last version, and it does not Oops, at least not for me.
The end result still looks odd to me:
[ 1.115219] scpi_clocks scpi:clocks: failed to register clock 'vcpu'
[ 1.159490] cpu cpu0: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 0, cluster: 0
[ 1.162986] cpu cpu0: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[ 1.170945] cpu cpu1: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 1, cluster: 0
[ 1.179634] cpu cpu1: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[ 1.187654] cpu cpu2: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 2, cluster: 0
[ 1.196284] cpu cpu2: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[ 1.204375] cpu cpu3: _get_cluster_clk_and_freq_table: Failed to get clk for
cpu: 3, cluster: 0
[ 1.212911] cpu cpu3: _get_cluster_clk_and_freq_table: Failed to get data for
cluster: 0
[ 1.220612] arm_big_little: bL_cpufreq_register: Registered platform driver:
scpi
So now, I have an scpi clock provider which registers successfully but fails to
register its only clock. As a consequence, I also have a cpufreq driver which
manages to register but has no clock cpu clock to drive ...
Yes, I agree the above is not entirely acceptable situation.
--
Regards,
Sudeep
From: Stephen Boyd <hidden> Date: 2017-06-30 00:25:37
On 06/28, Jerome Brunet wrote:
Current implementation of scpi_clk_add just print a warning when clock
fails to register but then keep going as if nothing happened. The
provider is then registered with bogus data.
This may latter lead to an Oops in __clk_create_clk when
hlist_add_head(&clk->clks_node, &hw->core->clks) is called.
This patch fixes the issue and errors if a clock fails to register.
Fixes: cd52c2a4b5c4 ("clk: add support for clocks provided by SCP(System Control Processor)")
Signed-off-by: Jerome Brunet <jbrunet@baylibre.com>
---
Applied to clk-next
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project