[PATCH 2/3] arm64: Work around Falkor erratum 1003

[Catalin Marinas]

On Wed, Dec 07, 2016 at 03:00:26PM -0500, Christopher Covington wrote:
> From: Shanker Donthineni <shankerd at codeaurora.org>
> 
> On the Qualcomm Datacenter Technologies Falkor v1 CPU, memory accesses may
> allocate TLB entries using an incorrect ASID when TTBRx_EL1 is being
> updated. Changing the TTBRx_EL1[ASID] and TTBRx_EL1[BADDR] fields
> separately using a reserved ASID will ensure that there are no TLB entries
> with incorrect ASID after changing the the ASID.
> 
> Pseudo code:
>   write TTBRx_EL1[ASID] to a reserved value
>   ISB
>   write TTBRx_EL1[BADDR] to a desired value
>   ISB
>   write TTBRx_EL1[ASID] to a desired value
>   ISB

While the new ASID probably won't have incorrect TLB entries, the
reserved ASID will have random entries from all over the place. That's
because in step 1 you change the ASID to the reserved one while leaving
the old BADDR in place. There is a brief time before changing the ASID
when speculative page table walks will populate the TLB with entries
tagged with the reserved ASID. Such entries are never removed during TLB
shoot-down for the real ASID, so, depending on how this CPU implements
the walk cache, you could end up with intermediate level entries still
active and pointing to freed/reused pages. It will eventually hit an
entry that looks global with weird consequences.

We've been bitten by this in the past on arm32: 52af9c6cd863 ("ARM:
6943/1: mm: use TTBR1 instead of reserved context ID").

-- 
Catalin