Re: [PATCH v3 3/5] arch/x86/mm: Refactor cond_ibpb() to support other use cases
From: Thomas Gleixner
Date: Sat Apr 18 2020 - 06:00:05 EST
"Singh, Balbir" <sblbir@xxxxxxxxxx> writes:
> On Fri, 2020-04-17 at 15:07 +0200, Thomas Gleixner wrote:
>>
>> Balbir Singh <sblbir@xxxxxxxxxx> writes:
>> >
>> > /*
>> > - * Use bit 0 to mangle the TIF_SPEC_IB state into the mm pointer which is
>> > - * stored in cpu_tlb_state.last_user_mm_ibpb.
>> > + * Bits to mangle the TIF_SPEC_IB state into the mm pointer which is
>> > + * stored in cpu_tlb_state.last_user_mm_spec.
>> > */
>> > #define LAST_USER_MM_IBPB 0x1UL
>> > +#define LAST_USER_MM_SPEC_MASK (LAST_USER_MM_IBPB)
>> >
>> > /* Reinitialize tlbstate. */
>> > - this_cpu_write(cpu_tlbstate.last_user_mm_ibpb, LAST_USER_MM_IBPB);
>> > + this_cpu_write(cpu_tlbstate.last_user_mm_spec, LAST_USER_MM_IBPB);
>>
>> Shouldn't that be LAST_USER_MM_MASK?
>>
> No, that crashes the system for SW flushes, because it tries to flush the L1D
> via the software loop and early enough we don't have the l1d_flush_pages
> allocated. LAST_USER_MM_MASK has LAST_USER_MM_FLUSH_L1D bit set.
You can trivially prevent this by checking l1d_flush_pages != NULL.
Thanks,
tglx