LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Robin Murphy <robin.murphy@arm.com>
To: Yisheng Xie <xieyisheng1@huawei.com>,
	will.deacon@arm.com, joro@8bytes.org
Cc: linux-arm-kernel@lists.infradead.org,
	iommu@lists.linux-foundation.org, linux-kernel@vger.kernel.org,
	liubo95@huawei.com, dingtianhong@huawei.com
Subject: Re: [RFC 2/2] iommu/arm-smmu-v3: Support software retention for pm_resume
Date: Mon, 23 Apr 2018 17:14:45 +0100	[thread overview]
Message-ID: <f947a71e-4348-c4f7-d900-9b75ca5bc073@arm.com> (raw)
In-Reply-To: <1524483959-54940-3-git-send-email-xieyisheng1@huawei.com>

On 23/04/18 12:45, Yisheng Xie wrote:
> When system suspend, hisilicon's smmu will do power gating for smmu,
> this time smmu's reg will be set to default value for not having
> hardware retention, which means need software do the retention instead.
> 
> The patch is to use arm_smmu_device_reset() to restore the register of
> smmu. However, it need to save the msis setting at probe if smmu do not
> support hardware retention.
> 
> Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
> ---
>   drivers/iommu/arm-smmu-v3.c | 69 +++++++++++++++++++++++++++++++++++++++++++--
>   1 file changed, 66 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
> index 044df6e..6cb56d8 100644
> --- a/drivers/iommu/arm-smmu-v3.c
> +++ b/drivers/iommu/arm-smmu-v3.c
> @@ -534,6 +534,11 @@ struct arm_smmu_strtab_cfg {
>   	u32				strtab_base_cfg;
>   };
>   
> +struct arm_smmu_msi_val {
> +	u64				doorbell;
> +	u32				data;
> +};

What does this do that struct msi_msg doesn't already (apart from take 
up more space in an array)?

> +
>   /* An SMMUv3 instance */
>   struct arm_smmu_device {
>   	struct device			*dev;
> @@ -558,6 +563,7 @@ struct arm_smmu_device {
>   
>   #define ARM_SMMU_OPT_SKIP_PREFETCH	(1 << 0)
>   #define ARM_SMMU_OPT_PAGE0_REGS_ONLY	(1 << 1)
> +#define ARM_SMMU_OPT_SW_RETENTION	(1 << 2)
>   	u32				options;
>   
>   	struct arm_smmu_cmdq		cmdq;
> @@ -587,6 +593,8 @@ struct arm_smmu_device {
>   
>   	u32				sync_count;
>   
> +	struct arm_smmu_msi_val		*msi;
> +	bool				probed;

This looks really hacky. I'm sure there's probably enough driver model 
information to be able to identify the probe state from just the struct 
device, but that's still not the right way to go. If you need to know 
this, then it can only mean we've got one-time software state 
initialisation mixed in with the actual hardware reset which programs 
the software state into the device. Thus there should be some 
refactoring to properly separate those concerns.

>   	bool				bypass;
>   
>   	/* IOMMU core code handle */
> @@ -630,6 +638,7 @@ struct arm_smmu_option_prop {
>   static struct arm_smmu_option_prop arm_smmu_options[] = {
>   	{ ARM_SMMU_OPT_SKIP_PREFETCH, "hisilicon,broken-prefetch-cmd" },
>   	{ ARM_SMMU_OPT_PAGE0_REGS_ONLY, "cavium,cn9900-broken-page1-regspace"},
> +	{ ARM_SMMU_OPT_SW_RETENTION, "hisilicon,broken-hardware-retention" },

That seems a bit over-specific - there are going to be any number of 
SMMU implementations/integrations which may or may not implement 
hardware retention states. More crucially, it's also backwards. Making 
the driver assume that *every* SMMU implements hardware retention unless 
this new DT property is present is quite obviously completely wrong, 
especially for ACPI...

The sensible thing to do is to implement suspend/resume support which 
works in general, *then* consider optimising it for cases where 
explicitly restoring the hardware state may be skipped (if indeed it 
makes a significant difference). Are there not already generic DT/ACPI 
properties for describing the retention levels of different power 
states, which could be made use of here?

>   	{ 0, NULL},
>   };
>   
> @@ -2228,7 +2237,8 @@ static void arm_smmu_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
>   	phys_addr_t doorbell;
>   	struct device *dev = msi_desc_to_dev(desc);
>   	struct arm_smmu_device *smmu = dev_get_drvdata(dev);
> -	phys_addr_t *cfg = arm_smmu_msi_cfg[desc->platform.msi_index];
> +	int msi_index = desc->platform.msi_index;
> +	phys_addr_t *cfg = arm_smmu_msi_cfg[msi_index];
>   
>   	doorbell = (((u64)msg->address_hi) << 32) | msg->address_lo;
>   	doorbell &= MSI_CFG0_ADDR_MASK;
> @@ -2236,6 +2246,28 @@ static void arm_smmu_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
>   	writeq_relaxed(doorbell, smmu->base + cfg[0]);
>   	writel_relaxed(msg->data, smmu->base + cfg[1]);
>   	writel_relaxed(ARM_SMMU_MEMATTR_DEVICE_nGnRE, smmu->base + cfg[2]);
> +
> +	if (smmu->options & ARM_SMMU_OPT_SW_RETENTION) {

The overhead of writing an extra 12 bytes per MSI to memory is entirely 
negligible; saving the message data just doesn't warrant the complexity 
of being conditional. In fact, given the need to untangle the IRQ 
requests from the hardware reset, I'd rather expect to end up *only* 
saving the message here, and writing the IRQ_CFG registers later along 
with everything else.

> +		smmu->msi[msi_index].doorbell = doorbell;
> +		smmu->msi[msi_index].data = msg->data;
> +	}
> +}
> +
> +static void arm_smmu_restore_msis(struct arm_smmu_device *smmu)
> +{
> +	int nevc = ARM_SMMU_MAX_MSIS - 1;
> +
> +	if (!(smmu->features & ARM_SMMU_FEAT_PRI))
> +		nevc--;
> +
> +	for (; nevc >= 0; nevc--) {
> +		phys_addr_t *cfg = arm_smmu_msi_cfg[nevc];
> +		struct arm_smmu_msi_val msi_val = smmu->msi[nevc];
> +
> +		writeq_relaxed(msi_val.doorbell, smmu->base + cfg[0]);
> +		writel_relaxed(msi_val.data, smmu->base + cfg[1]);
> +		writel_relaxed(ARM_SMMU_MEMATTR_DEVICE_nGnRE, smmu->base + cfg[2]);
> +	}
>   }
>   
>   static void arm_smmu_setup_msis(struct arm_smmu_device *smmu)
> @@ -2261,6 +2293,16 @@ static void arm_smmu_setup_msis(struct arm_smmu_device *smmu)
>   		return;
>   	}
>   
> +	if (smmu->probed) {
> +		BUG_ON(!(smmu->options & ARM_SMMU_OPT_SW_RETENTION));
> +		arm_smmu_restore_msis(smmu);
> +		return;
> +	} else if (smmu->options & ARM_SMMU_OPT_SW_RETENTION) {
> +		smmu->msi = devm_kmalloc_array(dev, nvec,
> +					       sizeof(*(smmu->msi)),
> +					       GFP_KERNEL);
> +	}

A single code path which either allocates an empty array *or* writes the 
contents of that array to hardware is a clear "you're doing it wrong" 
indicator. Yes, this definitely wants refactoring.

> +
>   	/* Allocate MSIs for evtq, gerror and priq. Ignore cmdq */
>   	ret = platform_msi_domain_alloc_irqs(dev, nvec, arm_smmu_write_msi_msg);
>   	if (ret) {
> @@ -2294,6 +2336,9 @@ static void arm_smmu_setup_unique_irqs(struct arm_smmu_device *smmu)
>   
>   	arm_smmu_setup_msis(smmu);
>   
> +	if (smmu->probed)
> +		return;
> +
>   	/* Request interrupt lines */
>   	irq = smmu->evtq.q.irq;
>   	if (irq) {
> @@ -2348,7 +2393,7 @@ static int arm_smmu_setup_irqs(struct arm_smmu_device *smmu)
>   	}
>   
>   	irq = smmu->combined_irq;
> -	if (irq) {
> +	if (irq && !smmu->probed) {
>   		/*
>   		 * Cavium ThunderX2 implementation doesn't not support unique
>   		 * irq lines. Use single irq line for all the SMMUv3 interrupts.
> @@ -2360,7 +2405,7 @@ static int arm_smmu_setup_irqs(struct arm_smmu_device *smmu)
>   					"arm-smmu-v3-combined-irq", smmu);
>   		if (ret < 0)
>   			dev_warn(smmu->dev, "failed to enable combined irq\n");
> -	} else
> +	} else if (!irq)
>   		arm_smmu_setup_unique_irqs(smmu);
>   
>   	if (smmu->features & ARM_SMMU_FEAT_PRI)
> @@ -2882,6 +2927,9 @@ static int arm_smmu_device_probe(struct platform_device *pdev)
>   		if (ret)
>   			return ret;
>   	}
> +
> +	smmu->probed = true;
> +
>   	return 0;
>   }
>   
> @@ -2899,6 +2947,20 @@ static void arm_smmu_device_shutdown(struct platform_device *pdev)
>   	arm_smmu_device_remove(pdev);
>   }
>   
> +static int arm_smmu_pm_resume(struct device *dev)
> +{
> +	struct arm_smmu_device *smmu = dev_get_drvdata(dev);
> +
> +	if (smmu->options & ARM_SMMU_OPT_SW_RETENTION)
> +		return arm_smmu_device_reset(smmu);

Given the SYSTEM_SLEEP_PM_OPS below, does your hardware really preserve 
state across hibernate/restore as well? That would be particularly 
impressive ;)

Robin.

> +
> +	return 0;
> +}
> +
> +static const struct dev_pm_ops arm_smmu_pm_ops = {
> +	SET_SYSTEM_SLEEP_PM_OPS(NULL, arm_smmu_pm_resume)
> +};
> +
>   static const struct of_device_id arm_smmu_of_match[] = {
>   	{ .compatible = "arm,smmu-v3", },
>   	{ },
> @@ -2909,6 +2971,7 @@ static void arm_smmu_device_shutdown(struct platform_device *pdev)
>   	.driver	= {
>   		.name		= "arm-smmu-v3",
>   		.of_match_table	= of_match_ptr(arm_smmu_of_match),
> +		.pm		= &arm_smmu_pm_ops,
>   	},
>   	.probe	= arm_smmu_device_probe,
>   	.remove	= arm_smmu_device_remove,
> 

  reply	other threads:[~2018-04-23 16:15 UTC|newest]

Thread overview: 7+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2018-04-23 11:45 [RFC 0/2] " Yisheng Xie
2018-04-23 11:45 ` [RFC 1/2] iommu/arm-smmu-v3: Remove bypass in arm_smmu_reset_device Yisheng Xie
2018-04-23 16:07   ` Robin Murphy
2018-04-24 11:57     ` Yisheng Xie
2018-04-23 11:45 ` [RFC 2/2] iommu/arm-smmu-v3: Support software retention for pm_resume Yisheng Xie
2018-04-23 16:14   ` Robin Murphy [this message]
2018-04-24 11:56     ` Yisheng Xie

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=f947a71e-4348-c4f7-d900-9b75ca5bc073@arm.com \
    --to=robin.murphy@arm.com \
    --cc=dingtianhong@huawei.com \
    --cc=iommu@lists.linux-foundation.org \
    --cc=joro@8bytes.org \
    --cc=linux-arm-kernel@lists.infradead.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=liubo95@huawei.com \
    --cc=will.deacon@arm.com \
    --cc=xieyisheng1@huawei.com \
    --subject='Re: [RFC 2/2] iommu/arm-smmu-v3: Support software retention for pm_resume' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).