[PATCH 3/7] powerpc/64s: flush L1D after user accesses

[Christophe Leroy]

Quoting Qian Cai <qcai at redhat.com>:

> On Thu, 2020-12-03 at 12:17 -0500, Qian Cai wrote:
>> []
>> > +static inline bool
>> > +bad_kuap_fault(struct pt_regs *regs, unsigned long address, bool  
>> is_write)
>> > +{
>> > +	return WARN(mmu_has_feature(MMU_FTR_RADIX_KUAP) &&
>> > +		    (regs->kuap & (is_write ? AMR_KUAP_BLOCK_WRITE :  
>> AMR_KUAP_BLOCK_READ)),
>> > +		    "Bug: %s fault blocked by AMR!", is_write ? "Write" : "Read");
>> > +}
>>
>> A simple "echo t > /proc/sysrq-trigger" will trigger this warning almost
>> endlessly on POWER9 NV.
>
> I have just realized the patch just moved this warning around, so  
> the issue was
> pre-existent. Since I have not tested sysrq-t regularly, I am not  
> sure when it
> started to break. So far, I have reverted some of those for testing which did
> not help, i.e., the sysrq-t issue remains.
>
> 16852975f0f  Revert "powerpc/64s: Use early_mmu_has_feature() in set_kuap()"
> 129e240ead32 Revert "powerpc: Implement user_access_save() and  
> user_access_restore()"
> edb0046c842c Revert "powerpc/64s/kuap: Add missing isync to KUAP  
> restore paths"
> 2d46ee87ce44 Revert "powerpc/64/kuap: Conditionally restore AMR in  
> interrupt exit"
> c1e0e805fc57 Revert "powerpc/64s/kuap: Conditionally restore AMR in  
> kuap_restore_amr asm"
> 7f30b7aaf23a Revert "selftests/powerpc: rfi_flush: disable entry  
> flush if present"
> bc9b9967a100 Revert "powerpc/64s: flush L1D on kernel entry"
> b77e7b54f5eb Revert "powerpc/64s: flush L1D after user accesses"
> 22dddf532c64 Revert "powerpc: Only include kup-radix.h for 64-bit Book3S"
> 2679d155c46a Revert "selftests/powerpc: entry flush test"
> 87954b9b4243 Revert "selftests/powerpc: refactor entry and rfi_flush tests"
> 342d82bd4c5d Revert "powerpc/64s: rename pnv|pseries_setup_rfi_flush  
> to _setup_security_mitigations"

I also hit that WARNING in the same way earlier this week.

I think it has been broken by commit c33165253492 ("powerpc: use  
non-set_fs based maccess routines")

IIUC we should provide copy_from_kernel_nofault_allowed() to avoid that.

Christophe