linux-kernel - Re: [PATCH v3 1/1] iommu/arm-smmu-v3: Simplify useless instructions in arm_smmu_cmdq_build

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <4f296b55-a59a-49b4-ad4d-902bb83a203c@huawei.com>
Date:   Wed, 8 Dec 2021 18:17:45 +0000
From:   John Garry <john.garry@...wei.com>
To:     "Leizhen (ThunderTown)" <thunder.leizhen@...wei.com>,
        Will Deacon <will@...nel.org>,
        Robin Murphy <robin.murphy@....com>,
        Joerg Roedel <joro@...tes.org>,
        <linux-arm-kernel@...ts.infradead.org>,
        <iommu@...ts.linux-foundation.org>, <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v3 1/1] iommu/arm-smmu-v3: Simplify useless instructions
 in arm_smmu_cmdq_build_cmd()

>> Did you notice any performance change with this change?
> 
> Hi John:
>    Thanks for the tip. I wrote a test case today, and I found that the
> performance did not go up but down.

I very quickly tested on a DMA mapping benchmark very similar to the 
kernel DMA benchmark module - I got mixed results. For fewer CPUs (<8), 
a small improvement, like 0.7%. For more CPUs, a dis-improvement - 
that's surprising, I did expect just no change as any improvement would 
get dwarfed from the slower unmap rates for more CPUs. I can check this
more tomorrow.

> It's so weird. So I decided not to
> change it, because it's also poorly readable. So I plan to make only
> the following modifications:
> @@ -237,7 +237,7 @@ static int queue_remove_raw(struct arm_smmu_queue *q, u64 *ent)
>   static int arm_smmu_cmdq_build_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
>   {
>          memset(cmd, 0, 1 << CMDQ_ENT_SZ_SHIFT);
> -       cmd[0] |= FIELD_PREP(CMDQ_0_OP, ent->opcode);
> +       cmd[0] = FIELD_PREP(CMDQ_0_OP, ent->opcode);
> 
>          switch (ent->opcode) {
>          case CMDQ_OP_TLBI_EL2_ALL:
> 
> This prevents the compiler from generating the following two inefficient
> instructions:
>       394:       f9400002        ldr     x2, [x0]	//x2 = cmd[0]
>       398:       aa020062        orr     x2, x3, x2	//x3 = FIELD_PREP(CMDQ_0_OP, ent->opcode)
> 
> Maybe it's not worth changing because I've only seen a 0.x nanosecond reduction
> in performance. But one thing is, it only comes with benefits, no side effects.
> 

I just think that with the original code that cmd[] is on the stack and 
cached, so if we have write-back attribute (which I think we do) then 
there would not necessarily a write to external memory per write to cmd[].

So, apart from this approach, I think that if we can just reduce the 
instructions through other efficiencies in the function then that would 
be good.

Thanks,
John