On 9/27/18 10:29 PM, Youness Alaoui wrote:
Thanks everyone for the responses. So far my understanding on the task at hand is :
- Add a CONFIG to decide if we set FLOCKDN or not (and one to decide
if we lock it on the resume path?)
Maybe no config at all, see discussion of PRR34_LOCKDN below. But if, then only one config, IMO: On the resume path FLOCKDN should always be set.
- Remove the chipset_lockdown devicetree config and change the code to
always assume it's LOCKDOWN_COREBOOT (this applies to all platforms, right?)
Yes, unless somebody objects. Please make sure to add the right reviewers to this patch. I expect some resistance.
- remove the call to fast_spi_pr_dlock() and fix the comment above
that function to be more accurate
Please also check the tree for code that sets PRRs. IIRC, it is used somewhere to protect the MRC cache. I would set the discrete lock directly after writing the region.
- Add .config CONFIG options or maybe optional devicetree configs for
how to setup the registers so coreboot can prepare them (due to the resume path) ?
Hmm, I would add Kconfig options for this as well. I you want to match the configuration of your payload, it definitely doesn't belong into the devicetree. Note that "prepare" also implies to lock them on the resume path.
AIUI, FLOCKDN always locks the PRRs.
Actually from what I can see, the FLOCKDN will lock a few registers, among which it will lock the "discrete PR locks" register, and it will lock the PR 0, 1 and 2, but there's also a PRR34_LOCKDN bit to separately lock the PRR 3 and 4. So technically, I could leave FLOCKDN set, and in the payload, just set protected range in PRR3 or PRR4, then set the PRR34_LOCKDN. Unfortunately, I can't do that because the PRR3 and 4 were already locked with the discrete lock register (if I remove the call to fast_spi_pr_dlock(), it should be fine for my needs though I think, since you said swseq would also fail for protected ranges).
Ah, that PRR34_LOCKDN was only introduced with SKL. Didn't know it or at least didn't remember it. In this case it seems best to always set FLOCKDN in coreboot. The payload can then still protect additional regions with PRR3/4. On the resume path, coreboot would set both FLOCKDN and PRR34_LOCKDN, right?
Also note that swseq is not an option anymore since Skylake (IIRC, the ME protects the flash-cycle go bit in its default configuration). With only hwseq, you can only access the flash chip's first status register which makes many security features unusable. So we have to rely on the protected regions :-/ I hope Intel is going to fix that for future plat- forms (or with an ME update if that could add more hwseq commands).
Well, I saw the opmenu/optypes/etc.. registers still being set in skylake (their offsets were not in the datasheet though), so I figured swseq is still there, just undocumented. It might work on systems where the ME is disabled. I'd have to test. I was thinking of using swseq and add the write-register-2 opcode and use it to set SPR1 and lock the entire flash until it gets power cycled. I don't know if that would work during suspend, but if it does, it might help with the resume path at least (but would prevent updating flash after a reboot and would force user to shut down and boot). I guess I have some trial-error to do.
The tricky part is to get access to the flash chip's second status register. I couldn't figure any way to do that with SKL (this cost me some weeks of work once because FSP also locked empty flash pro- tection).
Actually, if status register remains the same through a reboot/suspend/resume, and swseq doesn't work on skylake at all, I could just set the protect bits and prevent writing to status register on resume path.
I don't understand, is there a lock to prevent status register writes with hwseq? Also I though, you can't disable these status register bits w/o a reset. Whether the flash chip is reset, depends on your board design and power sequencing ofc.
One way would be to let coreboot decide, e.g. prepare the configuration and don't lock it, and let the payload lock. The payload could validate this configuration before locking (and issue a warning if coreboot didn't set the expected bits).
Thanks, that's a good idea, I forgot about the suspend/resume case. This implies I need coreboot to know about what config the payload will want to set. I'm not sure if that should go into a CONFIG or a devicetree config. Since it might be user-configured, it makes more sense for it to go into .config, right ?
Yes, Kconfig seems to be the right place.
Nico
On Sat, Sep 29, 2018 at 12:21 PM Nico Huber nico.h@gmx.de wrote:
On 9/27/18 10:29 PM, Youness Alaoui wrote:
Thanks everyone for the responses. So far my understanding on the task at hand is :
- Add a CONFIG to decide if we set FLOCKDN or not (and one to decide
if we lock it on the resume path?)
Maybe no config at all, see discussion of PRR34_LOCKDN below. But if, then only one config, IMO: On the resume path FLOCKDN should always be set.
- Remove the chipset_lockdown devicetree config and change the code to
always assume it's LOCKDOWN_COREBOOT (this applies to all platforms, right?)
Yes, unless somebody objects. Please make sure to add the right reviewers to this patch. I expect some resistance.
Ok, this is already done, but I want to finish/test my work before I send for review and I might not have time until the end of the week.
- remove the call to fast_spi_pr_dlock() and fix the comment above
that function to be more accurate
Please also check the tree for code that sets PRRs. IIRC, it is used somewhere to protect the MRC cache. I would set the discrete lock directly after writing the region.
Ok, thanks, will do. Note: I've also changed the fast_spi_pr_dlock() function to take 5 u8 arguments to allow the caller to lock each prr individually instead of being an all or nothing.
- Add .config CONFIG options or maybe optional devicetree configs for
how to setup the registers so coreboot can prepare them (due to the resume path) ?
Hmm, I would add Kconfig options for this as well. I you want to match the configuration of your payload, it definitely doesn't belong into the devicetree. Note that "prepare" also implies to lock them on the resume path.
Yeah, so far, the KConfigs I added are : FAST_SPI_LOCK_REGISTERS FAST_SPI_LOCK_REGISTERS_ON_RESUME FAST_SPI_LOCK_PRR34_REGISTERS FAST_SPI_LOCK_PRR34_REGISTERS_ON_RESUME FAST_SPI_PRR3_REGISTER (hex value) FAST_SPI_PRR4_REGISTER (hex value)
This way, a user can lock/unlock the register (since it affects more than just the PRRs) and they can override/set the PRR3 and PRR4 values and decide whether to lock them or not or only on resume, etc.. I think those should be enough for most use cases and it's enough for what I need.
AIUI, FLOCKDN always locks the PRRs.
Actually from what I can see, the FLOCKDN will lock a few registers, among which it will lock the "discrete PR locks" register, and it will lock the PR 0, 1 and 2, but there's also a PRR34_LOCKDN bit to separately lock the PRR 3 and 4. So technically, I could leave FLOCKDN set, and in the payload, just set protected range in PRR3 or PRR4, then set the PRR34_LOCKDN. Unfortunately, I can't do that because the PRR3 and 4 were already locked with the discrete lock register (if I remove the call to fast_spi_pr_dlock(), it should be fine for my needs though I think, since you said swseq would also fail for protected ranges).
Ah, that PRR34_LOCKDN was only introduced with SKL. Didn't know it or at least didn't remember it. In this case it seems best to always set FLOCKDN in coreboot. The payload can then still protect additional regions with PRR3/4. On the resume path, coreboot would set both FLOCKDN and PRR34_LOCKDN, right?
On resume, even if coreboot sets them, the PRRs need to be filled, which is why I also added the hex value of the PRR in the KConfig. I also want to keep FLOCKDN unlocked, so Heads can set the write protect bits in the flash then prevent writing to the status register via hwseq before locking it with FLOCKDN.
Also note that swseq is not an option anymore since Skylake (IIRC, the ME protects the flash-cycle go bit in its default configuration). With only hwseq, you can only access the flash chip's first status register which makes many security features unusable. So we have to rely on the protected regions :-/ I hope Intel is going to fix that for future plat- forms (or with an ME update if that could add more hwseq commands).
Well, I saw the opmenu/optypes/etc.. registers still being set in skylake (their offsets were not in the datasheet though), so I figured swseq is still there, just undocumented. It might work on systems where the ME is disabled. I'd have to test. I was thinking of using swseq and add the write-register-2 opcode and use it to set SPR1 and lock the entire flash until it gets power cycled. I don't know if that would work during suspend, but if it does, it might help with the resume path at least (but would prevent updating flash after a reboot and would force user to shut down and boot). I guess I have some trial-error to do.
The tricky part is to get access to the flash chip's second status register. I couldn't figure any way to do that with SKL (this cost me some weeks of work once because FSP also locked empty flash pro- tection).
Only way I see is through swseq, but you say it doesn't work because of the ME. I'd have to test if the disabled ME allows swseq to work.
Actually, if status register remains the same through a reboot/suspend/resume, and swseq doesn't work on skylake at all, I could just set the protect bits and prevent writing to status register on resume path.
I don't understand, is there a lock to prevent status register writes with hwseq? Also I though, you can't disable these status register bits w/o a reset. Whether the flash chip is reset, depends on your board design and power sequencing ofc.
Yes, see bit 11 of HSFSC register (WRSDIS). That one already has a Kconfig to decide if we want it locked or not.
One way would be to let coreboot decide, e.g. prepare the configuration and don't lock it, and let the payload lock. The payload could validate this configuration before locking (and issue a warning if coreboot didn't set the expected bits).
Thanks, that's a good idea, I forgot about the suspend/resume case. This implies I need coreboot to know about what config the payload will want to set. I'm not sure if that should go into a CONFIG or a devicetree config. Since it might be user-configured, it makes more sense for it to go into .config, right ?
Yes, Kconfig seems to be the right place.
Nico
Thanks!
-
Nico Huber -
Youness Alaoui