Re: [PATCH 5/6] reset: rockchip: secure reset must be used by SCMI

From: Sebastian Reichel
Date: Sun Nov 12 2023 - 15:26:48 EST

Next message: Paul Moore: "Re: [PATCH v4 4/6] add statmount(2) syscall"
Previous message: Helge Deller: "Re: Bisected stability regression in 6.6"
In reply to: Krzysztof Kozlowski: "Re: [PATCH 5/6] reset: rockchip: secure reset must be used by SCMI"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hi,

On Sat, Nov 11, 2023 at 10:28:59PM +0100, Krzysztof Kozlowski wrote:
> On 11/11/2023 21:51, Sebastian Reichel wrote:
> > Hi,
> >
> > On Tue, Nov 07, 2023 at 06:45:03PM +0100, Krzysztof Kozlowski wrote:
> >> On 07/11/2023 18:35, Heiko Stübner wrote:
> >>> Am Dienstag, 7. November 2023, 17:21:41 CET schrieb Krzysztof Kozlowski:
> >>>> On 07/11/2023 16:55, Corentin Labbe wrote:
> >>>>> While working on the rk3588 crypto driver, I loose lot of time
> >>>>> understanding why resetting the IP failed.
> >>>>> This is due to RK3588_SECURECRU_RESET_OFFSET being in the secure world,
> >>>>> so impossible to operate on it from the kernel.
> >>>>> All resets in this block must be handled via SCMI call.
> >>>>>
> >>>>> Signed-off-by: Corentin Labbe <clabbe@xxxxxxxxxxxx>
> >>>>> ---
> >>>>> drivers/clk/rockchip/rst-rk3588.c | 42 ------------
> >>>>> .../dt-bindings/reset/rockchip,rk3588-cru.h | 68 +++++++++----------
> >>>>
> >>>> Please run scripts/checkpatch.pl and fix reported warnings. Some
> >>>> warnings can be ignored, but the code here looks like it needs a fix.
> >>>> Feel free to get in touch if the warning is not clear.
> >>>>
> >>>>> 2 files changed, 34 insertions(+), 76 deletions(-)
> >>>>>
> >>>>> diff --git a/drivers/clk/rockchip/rst-rk3588.c b/drivers/clk/rockchip/rst-rk3588.c
> >>>>> index e855bb8d5413..6556d9d3c7ab 100644
> >>>>> --- a/drivers/clk/rockchip/rst-rk3588.c
> >>>>> +++ b/drivers/clk/rockchip/rst-rk3588.c
> >>>>> @@ -16,9 +16,6 @@
> >>>>> /* 0xFD7C8000 + 0x0A00 */
> >>>>> #define RK3588_PHPTOPCRU_RESET_OFFSET(id, reg, bit) [id] = (0x8000*4 + reg * 16 + bit)
> >>>>>
> >>>>> -/* 0xFD7D0000 + 0x0A00 */
> >>>>> -#define RK3588_SECURECRU_RESET_OFFSET(id, reg, bit) [id] = (0x10000*4 + reg * 16 + bit)
> >>>>> -
> >>>>> /* 0xFD7F0000 + 0x0A00 */
> >>>>> #define RK3588_PMU1CRU_RESET_OFFSET(id, reg, bit) [id] = (0x30000*4 + reg * 16 + bit)
> >>>>>
> >>>>> @@ -806,45 +803,6 @@ static const int rk3588_register_offset[] = {
> >>>>> RK3588_PMU1CRU_RESET_OFFSET(SRST_P_PMU0IOC, 5, 4),
> >>>>> RK3588_PMU1CRU_RESET_OFFSET(SRST_P_GPIO0, 5, 5),
> >>>>> RK3588_PMU1CRU_RESET_OFFSET(SRST_GPIO0, 5, 6),
> >>>>> -
> >>>>> - /* SECURECRU_SOFTRST_CON00 */
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_A_SECURE_NS_BIU, 0, 10),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_SECURE_NS_BIU, 0, 11),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_A_SECURE_S_BIU, 0, 12),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_SECURE_S_BIU, 0, 13),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_P_SECURE_S_BIU, 0, 14),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_CRYPTO_CORE, 0, 15),
> >>>>> -
> >>>>> - /* SECURECRU_SOFTRST_CON01 */
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_CRYPTO_PKA, 1, 0),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_CRYPTO_RNG, 1, 1),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_A_CRYPTO, 1, 2),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_CRYPTO, 1, 3),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_KEYLADDER_CORE, 1, 9),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_KEYLADDER_RNG, 1, 10),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_A_KEYLADDER, 1, 11),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_KEYLADDER, 1, 12),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_P_OTPC_S, 1, 13),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_OTPC_S, 1, 14),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_WDT_S, 1, 15),
> >>>>> -
> >>>>> - /* SECURECRU_SOFTRST_CON02 */
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_T_WDT_S, 2, 0),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_BOOTROM, 2, 1),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_A_DCF, 2, 2),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_P_DCF, 2, 3),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_BOOTROM_NS, 2, 5),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_P_KEYLADDER, 2, 14),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_TRNG_S, 2, 15),
> >>>>> -
> >>>>> - /* SECURECRU_SOFTRST_CON03 */
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_TRNG_NS, 3, 0),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_D_SDMMC_BUFFER, 3, 1),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_SDMMC, 3, 2),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_H_SDMMC_BUFFER, 3, 3),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_SDMMC, 3, 4),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_P_TRNG_CHK, 3, 5),
> >>>>> - RK3588_SECURECRU_RESET_OFFSET(SRST_TRNG_S, 3, 6),
> >>>>> };
> >>>>>
> >>>>> void rk3588_rst_init(struct device_node *np, void __iomem *reg_base)
> >>>>> diff --git a/include/dt-bindings/reset/rockchip,rk3588-cru.h b/include/dt-bindings/reset/rockchip,rk3588-cru.h
> >>>>> index d4264db2a07f..c0d08ae78cd5 100644
> >>>>> --- a/include/dt-bindings/reset/rockchip,rk3588-cru.h
> >>>>> +++ b/include/dt-bindings/reset/rockchip,rk3588-cru.h
> >>>>> @@ -716,39 +716,39 @@
> >>>>> #define SRST_P_GPIO0 627
> >>>>> #define SRST_GPIO0 628
> >>>>>
> >>>>> -#define SRST_A_SECURE_NS_BIU 629
> >>>>> -#define SRST_H_SECURE_NS_BIU 630
> >>>>> -#define SRST_A_SECURE_S_BIU 631
> >>>>> -#define SRST_H_SECURE_S_BIU 632
> >>>>> -#define SRST_P_SECURE_S_BIU 633
> >>>>> -#define SRST_CRYPTO_CORE 634
> >>>>> -
> >>>>> -#define SRST_CRYPTO_PKA 635
> >>>>> -#define SRST_CRYPTO_RNG 636
> >>>>> -#define SRST_A_CRYPTO 637
> >>>>> -#define SRST_H_CRYPTO 638
> >>>>> -#define SRST_KEYLADDER_CORE 639
> >>>>> -#define SRST_KEYLADDER_RNG 640
> >>>>> -#define SRST_A_KEYLADDER 641
> >>>>> -#define SRST_H_KEYLADDER 642
> >>>>> -#define SRST_P_OTPC_S 643
> >>>>> -#define SRST_OTPC_S 644
> >>>>> -#define SRST_WDT_S 645
> >>>>> -
> >>>>> -#define SRST_T_WDT_S 646
> >>>>> -#define SRST_H_BOOTROM 647
> >>>>> -#define SRST_A_DCF 648
> >>>>> -#define SRST_P_DCF 649
> >>>>> -#define SRST_H_BOOTROM_NS 650
> >>>>> -#define SRST_P_KEYLADDER 651
> >>>>> -#define SRST_H_TRNG_S 652
> >>>>> -
> >>>>> -#define SRST_H_TRNG_NS 653
> >>>>> -#define SRST_D_SDMMC_BUFFER 654
> >>>>> -#define SRST_H_SDMMC 655
> >>>>> -#define SRST_H_SDMMC_BUFFER 656
> >>>>> -#define SRST_SDMMC 657
> >>>>> -#define SRST_P_TRNG_CHK 658
> >>>>> -#define SRST_TRNG_S 659
> >>>>> +#define SRST_A_SECURE_NS_BIU 10
> >>>>
> >>>> NAK. You just broke all users.
> >>>
> >>> If I'm reading the commit message correctly, all resets in that area
> >>> couldn't have any users to begin with, as the registers controlling them
> >>> are in the secure space, and need a higher exception level
> >>>
> >>> So if anything is trying to handle these resets, would end up with some
> >>> security exception right now.
> >>>
> >>> Though I guess we might want to use different names and not reuse the
> >>> existing ones. scmi clocks use a SCMI_CLK_* id scheme, so maybe SCMI_SRST_* ?
> >>
> >> I don't quite get what the patch wants to achieve. Why dropping driver
> >> support for given reset ID is connected with changing the value of
> >> binding for given reset?
> >>
> >> What is the point of this define SRST_A_SECURE_NS_BIU 10?
> >
> > This is about two different reset controllers. The IDs defined here
> > are used by the operating system to access the correct registers.
> > The kernel has a LUT from the ID to a register addresses, which is
> > something you asked for during upstreaming.
> >
> > The ID defined by Corentin is for reset control via SCMI firmware,
> > which has different number scheme than Linux. To me the suggestion
> > from Heiko looks sensible (i.e. create a new ID scheme and keep the
> > current one unchanged).
>
> So the binding is not for Linux but for FW? This should be explained in
> the commit msg.

No.

The current binding describes reset IDs, which are mapped by the
Linux driver to register offsets in the CRU (clock-reset-unit).
But accessing the crypto reset line directly from Linux (which
usually does not run in secure state) will fail. Accessing it
from correct security context with the current binding is fine
though. Considering we are sharing the bindings with e.g.
U-Boot, I suggest to keep the currently defined IDs.

But Corentin tries to get this running on Linux. For that he
needs to ask the (SCMI) firmware running in secure state to
please take care of the reset. The firmware is using different
reset IDs (apparently the ones used by downstream Linux, which
are derived from register offset).

In DT the difference looks like this (check the different phandles):

#define SRST_A_SECURE_NS_BIU 629
crypto-old {
// existing binding from Linux perspective
// reset via direct CRU access
// NOTE: permission denied
resets = <&cru SRST_A_SECURE_NS_BIU>;
};

#define SCMI_RST_A_SECURE_NS_BIU 10
crypto-new {
// new binding from Linux perspective
// reset via SCMI firmware request
resets = <&scmi SCMI_RST_A_SECURE_NS_BIU>;
};

Instead of introducing SCMI_RST_A_SECURE_NS_BIU, Corentin
currently just redefines SRST_A_SECURE_NS_BIU. This is quite
misleading. If somebody does '<&cru SRST_A_SECURE_NS_BIU>'
with the '10' value for SCMI, it instead resets
SRST_A_TOP_M300_BIU.

So my suggestion is to go with the suggestion from Heiko and
introduce SCMI_RST_A_SECURE_NS_BIU (or something similar).
That also matches how the SCMI clks on RK3588 and some other
platforms. See e.g.:

of include/dt-bindings/clock/rockchip,rk3588-cru.h.

Greetings,

-- Sebastian

Next message: Paul Moore: "Re: [PATCH v4 4/6] add statmount(2) syscall"
Previous message: Helge Deller: "Re: Bisected stability regression in 6.6"
In reply to: Krzysztof Kozlowski: "Re: [PATCH 5/6] reset: rockchip: secure reset must be used by SCMI"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]