Re: [patches] RE: [PATCH 3/4] RISC-V: Flush I$ when making a dirty page executable

From: Andrew Waterman
Date: Wed Nov 22 2017 - 14:03:56 EST

On Wed, Nov 22, 2017 at 9:38 AM, Palmer Dabbelt <palmer@xxxxxxxxxx> wrote:
> On Tue, 21 Nov 2017 08:57:07 PST (-0800), David.Laight@xxxxxxxxxx wrote:
>> From: Palmer Dabbelt
>>> Sent: 20 November 2017 18:58
>>> The RISC-V ISA allows for instruction caches that are not coherent WRT
>>> stores, even on a single hart. As a result, we need to explicitly flush
>>> the instruction cache whenever marking a dirty page as executable in
>>> order to preserve the correct system behavior.
>> Isn't the I-flush only needed if there has been an unmap since the
>> previous I-flush?
>> Since code is rarely unmapped (exec and driver unload come to mind)
>> the I-flush won't be needed very often.
> There's nothing in the RISC-V ISA that prevents the instruction cache from
> caching read-only (or even unmapped!) pages. Instructions fetched this
> manner could never commit, but they could fill up the icache with garbage.
> I believe that means we need to flush on dirty->execute, but if I'm wrong
> I'm happy to change it.

Palmer is right about the ISA semantics; a FENCE.I needs to be
executed before executing any code that has been (physically)
modified, because any physical address could reside stale in the I$,
regardless of past virtual-memory permissions.

The approach we took to handle this (executing FENCE.I when a
phsyically dirty page becomes executable) is also used by the arm64