On 13.12.2019 04:03, Stephen Boyd wrote:
> Quoting Sylwester Nawrocki (2019-12-12 03:53:17)
>> Cc: Stephen and Michael
>>
>> On 11/21/19 11:11, Marek Szyprowski wrote:
>>> All top clocks on G3D path has to be enabled all the time to allow proper
>>> G3D power domain operation. This is achieved by adding CRITICAL flag to
>>> "mout_sw_aclk_g3d" clock, what keeps this clock and all its parents
>>> enabled.
>>>
>>> This fixes following imprecise abort issue observed on Odroid XU3/XU4
>>> after enabling Panfrost driver by commit 1a5a85c56402 "ARM: dts: exynos:
>>> Add Mali/GPU node on Exynos5420 and enable it on Odroid XU3/4"):
>>>
>>> panfrost 11800000.gpu: clock rate = 400000000
>>> panfrost 11800000.gpu: failed to get regulator: -517
>>> panfrost 11800000.gpu: regulator init failed -517
>>> Power domain G3D disable failed
>>> ...
>>> panfrost 11800000.gpu: clock rate = 400000000
>>> 8<--- cut here ---
>>> Unhandled fault: imprecise external abort (0x1406) at 0x00000000
>>> pgd = (ptrval)
>>> [00000000] *pgd=00000000
>>> Internal error: : 1406 [#1] PREEMPT SMP ARM
>>> Modules linked in:
>>> CPU: 7 PID: 53 Comm: kworker/7:1 Not tainted 5.4.0-rc8-next-20191119-00032-g56f1001191a6 #6923
>>> Hardware name: SAMSUNG EXYNOS (Flattened Device Tree)
>>> Workqueue: events deferred_probe_work_func
>>> PC is at panfrost_gpu_soft_reset+0x94/0x110
>>> LR is at ___might_sleep+0x128/0x2dc
>>> ...
>>> [<c05c231c>] (panfrost_gpu_soft_reset) from [<c05c2704>] (panfrost_gpu_init+0x10/0x67c)
>>> [<c05c2704>] (panfrost_gpu_init) from [<c05c15d0>] (panfrost_device_init+0x158/0x2cc)
>>> [<c05c15d0>] (panfrost_device_init) from [<c05c0cb0>] (panfrost_probe+0x80/0x178)
>>> [<c05c0cb0>] (panfrost_probe) from [<c05cfaa0>] (platform_drv_probe+0x48/0x9c)
>>> [<c05cfaa0>] (platform_drv_probe) from [<c05cd20c>] (really_probe+0x1c4/0x474)
>>> [<c05cd20c>] (really_probe) from [<c05cd694>] (driver_probe_device+0x78/0x1bc)
>>> [<c05cd694>] (driver_probe_device) from [<c05cb374>] (bus_for_each_drv+0x74/0xb8)
>>> [<c05cb374>] (bus_for_each_drv) from [<c05ccfa8>] (__device_attach+0xd4/0x16c)
>>> [<c05ccfa8>] (__device_attach) from [<c05cc110>] (bus_probe_device+0x88/0x90)
>>> [<c05cc110>] (bus_probe_device) from [<c05cc634>] (deferred_probe_work_func+0x4c/0xd0)
>>> [<c05cc634>] (deferred_probe_work_func) from [<c0149df0>] (process_one_work+0x300/0x864)
>>> [<c0149df0>] (process_one_work) from [<c014a3ac>] (worker_thread+0x58/0x5a0)
>>> [<c014a3ac>] (worker_thread) from [<c0151174>] (kthread+0x12c/0x160)
>>> [<c0151174>] (kthread) from [<c01010b4>] (ret_from_fork+0x14/0x20)
>>> Exception stack(0xee03dfb0 to 0xee03dff8)
>>> ...
>>> Code: e594300c e5933020 e3130c01 1a00000f (ebefff50).
>>> ---[ end trace badde2b74a65a540 ]---
>>>
>>> In the above case, the Panfrost driver disables G3D clocks after failure
>>> of getting the needed regulator and return with -EPROVE_DEFER code. This
>>> causes G3D power domain disable failure and then, during second probe
>>> an imprecise abort is triggered due to undefined power domain state.
>>>
>>> Fixes: 45f10dabb56b ("clk: samsung: exynos5420: Add SET_RATE_PARENT flag to clocks on G3D path")
>>> Fixes: c9f7567aff31 ("clk: samsung: exynos542x: Move G3D subsystem clocks to its sub-CMU")
>>> Signed-off-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx>
>>> ---
>>> drivers/clk/samsung/clk-exynos5420.c | 2 +-
>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/clk/samsung/clk-exynos5420.c b/drivers/clk/samsung/clk-exynos5420.c
>>> index 3a991ca1ee36..89126ba66995 100644
>>> --- a/drivers/clk/samsung/clk-exynos5420.c
>>> +++ b/drivers/clk/samsung/clk-exynos5420.c
>>> @@ -712,7 +712,7 @@ static const struct samsung_mux_clock exynos5x_mux_clks[] __initconst = {
>>> MUX(0, "mout_sw_aclk266_g2d", mout_sw_aclk266_g2d_p,
>>> SRC_TOP12, 12, 1),
>>> MUX_F(0, "mout_sw_aclk_g3d", mout_sw_aclk_g3d_p, SRC_TOP12, 16, 1,
>>> - CLK_SET_RATE_PARENT, 0),
>>> + CLK_IS_CRITICAL | CLK_SET_RATE_PARENT, 0),
>>> MUX(0, "mout_sw_aclk300_jpeg", mout_sw_aclk300_jpeg_p,
>> SRC_TOP12, 20, 1),
>>
>> Adding that flag to a mux clock doesn't look right, it feels we are not
>> addressing the issue properly and the root cause is not known.
>>
>> AFAICS CLK_IS_CRITICAL flag will also not ensure required root clock up
>> in the clk tree is enabled in case any reparenting is done after that mux
>> clock has been registered. The flag seems misused and the fix looks dubious
>> and fragile to me.
>>
>> I would apply that to fix crashes we are seeing in -next but we ned to come
>> up with a better solution.
>>
>> Perhaps we could add a comment like:
>>
>> "CLK_IS_CRITICAL flag is added to this clock as a workaround for imprecise
>> external abort triggered by the Panfrost driver and will be removed once
>> the issue is properly addressed."
> You should always add a comment to CLK_IS_CRITICAL usage. This sort of
> comment is not comforting though. When will the issue be properly
> addressed? Maybe we should block panfrost from being enabled on this
> platform until it is root caused?
I can add a comment, no problem.
This issue is not related to the panfrost driver at all. Panfrost driver
simply enables and disables the clock of its HW module. It is nothing
extraordinary.
The real issue here is that some SoC internal busses (not assigned to
any driver at all) are sourced from the same MUX, which that patch
flagged as CRITICAL without any additional gates, thus if the only
client of that MUX disables its gate clock, the whole path up to the
root PLL is disabled what causes the HW issue. The driver (or user via
dts) might change the root PLL, so that CRITICAL flag cannot be moved to
the top clock in this hierarchy.
I can change that CRITICAL flag to a explicit call to
clk_prepare_enable() during exynos542x-clk driver probe, but IMHO the
flag better fits in such case.
I'm not sure if the issue can be properly addressed in a better way. The
main problem here is that the whole clock topology for Exynos5422 SoCs
is not fully documented in datasheet (especially the relation of the
clocks and SoC's power domains) and the driver doesn't also fully model
it, because other Exynos542x HW module drivers doesn't expect such fine
grained clocks (mainly for historical compatibility).
I can take a look how to better model some problematic parts, but this
won't be a simple, single patch. The issue mitigated by this patch
already prevents v5.5-rc1 from booting with the default
exynos_defconfig/multi_v7_defconfig, so I would like to merge it as a
fix for v5.5 first.
Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
