On 11/19/19 2:59 PM, Andrew F. Davis wrote: > On 11/19/19 2:44 PM, Tony Lindgren wrote: >> * Andrew F. Davis <afd@xxxxxx> [191119 19:36]: >>> On 11/19/19 2:20 PM, Tony Lindgren wrote: >>>> * Andrew F. Davis <afd@xxxxxx> [191119 19:13]: >>>>> On 11/19/19 2:07 PM, Tony Lindgren wrote: >>>>>> * Andrew F. Davis <afd@xxxxxx> [191119 18:51]: >>>>>>> On 11/19/19 1:32 PM, Tony Lindgren wrote: >>>>>>>> It would allow us to completely change over to using >>>>>>>> arm_smccc_smc() and forget the custom calls. >>>>>>> >>>>>>> We would need more than just the r12 quirk to replace all our custom SMC >>>>>>> handlers, we would need quirks for omap_smc2 which puts process ID in r1 >>>>>>> and puts #0xff in r6, and omap_smc3 that uses smc #1. All of our legacy >>>>>>> SMC calls also trash r4-r11, that is very non SMCCC complaint as only >>>>>>> r4-r7 need be caller saved. I don't see arm_smccc_smc() working with >>>>>>> legacy ROM no matter how much we hack at it :( >>>>>> >>>>>> We would just have omap_smc2() call arm_smccc_smc() and in that >>>>>> case. And omap_smc2() would still deal with saving and restoring >>>>>> the registers. >>>>> >>>>> Then why call arm_smccc_smc()? omap_smc2() is already an assembly >>>>> function, all it needs to do after loading the registers and saving the >>>>> right ones is issue an "smc #0" instruction, why would we want to >>>>> instead call into some other function to re-save registers and issue the >>>>> exact same instruction? >>>> >>>> To use Linux generic API for smc calls where possible. >>> >>> But we are not using generic API calls, we are using omap_smcx() which >>> cannot call into arm_smccc_smc(). For all the above reasons plus >>> arm_smccc_smc() uses r12 to save the stack pointer, our ROM expects r12 >>> to store the function ID. >> >> Saving and restoring r12 could be handled by the arm_smccc_smc() quirk >> for the non-optee case. >> >> Then we could get rid of omap_smc1() and arm_smccc_smc() should work >> for the optee case and non-optee case, right. >> > > > Yes, we could have both cases working if we could get the quirk in. > > >>>>>> Certainly the wrapper functions calling arm_smccc_smc() can deal >>>>>> with r12 too if the r12-quirk version and the plain version are >>>>>> never needed the same time on a booted SoC. >>>>>> >>>>>> Are they ever needed the same time on a booted SoC or not? >> >>> They should not be needed at the same time, either OP-TEE is on the >>> secure side or ROM is there. >> >> OK thanks. So we could just modify the code dynamically on boot >> based on if optee is found or not. The quirk could be done along >> the lines of the qcom quirk but only for the non-optee case: >> > > > We wouldn't have to patch anything if we could get the quirk in. One has > to state they wish to use the quirk version in a structure passed into > arm_smccc_smc_quirk(), in which case for all legacy user we just fill > out this quirk struct. OP-TEE uses the same arm_smccc_smc() but without > the quirk struct and so it uses the compliant call. > > The issue is still the same, I tried adding this, I got NAKd, if you > want to convince Mark to change his mind and allow us the quirk then we > can go down this path. Otherwise this will remain a dead end. > Hi Tony, Looks like the TI quirk idea is not moving forward, even the QCOM quirk looks like it may get kicked out. arm_smccc_smc() will remain only for SMCCC compliant calls, but it looks like a generic arm_smc() wouldn't be too opposed upstream. Either way this patch would still be valid as when OP-TEE is present then arm_smccc_smc() will be the right call to make, how we handle the legacy calls can be sorted out later if a generic SMC call is implemented. Thanks, Andrew > Andrew > > >> $ git grep -C10 ARM_SMCCC_QUIRK_QCOM_A6 >> >> Regards, >> >> Tony >>
