> -----Original Message-----
> From: Robin Murphy [mailto:robin.mur...@arm.com]
> Sent: Friday, July 17, 2020 8:55 PM
> To: Song Bao Hua (Barry Song) <song.bao....@hisilicon.com>; w...@kernel.org;
> j...@8bytes.org
> Cc: linux-ker...@vger.kernel.org; Linuxarm <linux...@huawei.com>;
> linux-arm-ker...@lists.infradead.org; iommu@lists.linux-foundation.org;
> Zengtao (B) <prime.z...@hisilicon.com>
> Subject: Re: [PATCH] iommu/arm-smmu-v3: remove the approach of MSI
> polling for CMD SYNC
> On 2020-07-17 00:07, Barry Song wrote:
> > Before commit 587e6c10a7ce ("iommu/arm-smmu-v3: Reduce contention
> during
> > command-queue insertion"), msi polling perhaps performed better since
> > it could run outside the spin_lock_irqsave() while the code polling cons
> > reg was running in the lock.
> >
> > But after the great reorganization of smmu queue, neither of these two
> > polling methods are running in a spinlock. And real tests show polling
> > cons reg via sev means smaller latency. It is probably because polling
> > by msi will ask hardware to write memory but sev polling depends on the
> > update of register only.
> >
> > Using 16 threads to run netperf on hns3 100G NIC with UDP packet size
> > in 32768bytes and set iommu to strict, TX throughput can improve from
> > 25227.74Mbps to 27145.59Mbps by this patch. In this case, SMMU is super
> > busy as hns3 sends map/unmap requests extremely frequently.
> How many different systems and SMMU implementations are those numbers
> representative of? Given that we may have cases where the SMMU can use
> MSIs but can't use SEV, so would have to fall back to inefficient
> busy-polling, I'd be wary of removing this entirely. Allowing particular
> platforms or SMMU implementations to suppress MSI functionality if they
> know for sure it makes sense seems like a safer bet.
Hello Robin,

Thanks for taking a look. Actually I was really struggling with the good way to 
make every platform happy.
And I don't have other platforms to test and check if those platforms run 
better by sev polling. Even two
platforms have completely same SMMU features, it is still possible they behave 
So I simply sent this patch to get the discussion started to get opinions.

At the first beginning, I wanted to have a module parameter for users to decide 
if msi polling should be disabled.
But the module parameter might be totally ignored by linux distro.

--- a/drivers/iommu/arm-smmu-v3.c
+++ b/drivers/iommu/arm-smmu-v3.c
@@ -418,6 +418,11 @@ module_param_named(disable_bypass, disable_bypass, bool, 
S_IRUGO);  MODULE_PARM_DESC(disable_bypass,
        "Disable bypass streams such that incoming transactions from devices 
that are not attached to an iommu domain will report an abort back to the 
device and will not be allowed to pass through the SMMU.");
+static bool disable_msipolling = 1;
+module_param_named(disable_msipolling, disable_msipolling, bool, 
+S_IRUGO); MODULE_PARM_DESC(disable_msipolling,
+       "Don't use MSI to poll the completion of CMD_SYNC if it is slower than 
 enum pri_resp {
        PRI_RESP_DENY = 0,
        PRI_RESP_FAIL = 1,
@@ -992,7 +997,7 @@ static void arm_smmu_cmdq_build_sync_cmd(u64 *cmd, struct 
arm_smmu_device *smmu,
         * Beware that Hi16xx adds an extra 32 bits of goodness to its MSI
         * payload, so the write will zero the entire command on that platform.
-       if (smmu->features & ARM_SMMU_FEAT_MSI &&
+       if (!disable_msipolling && smmu->features & ARM_SMMU_FEAT_MSI &&
            smmu->features & ARM_SMMU_FEAT_COHERENCY) {
                ent.sync.msiaddr = q->base_dma + Q_IDX(&q->llq, prod) *
                                   q->ent_dwords * 8;
@@ -1332,7 +1337,7 @@ static int __arm_smmu_cmdq_poll_until_consumed(struct 
arm_smmu_device *smmu,  static int arm_smmu_cmdq_poll_until_sync(struct 
arm_smmu_device *smmu,
                                         struct arm_smmu_ll_queue *llq)
-       if (smmu->features & ARM_SMMU_FEAT_MSI &&
+       if (!disable_msipolling && smmu->features & ARM_SMMU_FEAT_MSI &&
            smmu->features & ARM_SMMU_FEAT_COHERENCY)
                return __arm_smmu_cmdq_poll_until_msi(smmu, llq);

Another option is that we don't use module parameter, alternatively, we check 
the vendor/chip ID,
if the chip has better performance on sev polling, it may set 
disable_msipolling to true.

You are very welcome to give your suggestions.

iommu mailing list

Reply via email to