linux-kernel - Re: [PATCH 3/7] iommu/arm-smmu: Add tlb

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: Windows password security audit tool. GUI, reports in PDF.

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

Message-ID: <554f8de1-1638-4eb9-59ae-8e1f0d786c44@arm.com>
Date:   Fri, 30 Aug 2019 16:23:13 +0100
From:   Robin Murphy <robin.murphy@....com>
To:     Krishna Reddy <vdumpa@...dia.com>
Cc:     snikam@...dia.com, thomasz@...dia.com, jtukkinen@...dia.com,
        mperttunen@...dia.com, praithatha@...dia.com,
        iommu@...ts.linux-foundation.org, linux-kernel@...r.kernel.org,
        talho@...dia.com, yhsu@...dia.com, linux-tegra@...r.kernel.org,
        treding@...dia.com, avanbrunt@...dia.com,
        linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH 3/7] iommu/arm-smmu: Add tlb_sync implementation hook

On 29/08/2019 23:47, Krishna Reddy wrote:
> tlb_sync hook allows nvidia smmu handle tlb sync
> across multiple SMMUs as necessary.
> 
> Signed-off-by: Krishna Reddy <vdumpa@...dia.com>
> ---
>   drivers/iommu/arm-smmu-nvidia.c | 32 ++++++++++++++++++++++++++++++++
>   drivers/iommu/arm-smmu.c        |  8 +++++---
>   drivers/iommu/arm-smmu.h        |  4 ++++
>   3 files changed, 41 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/iommu/arm-smmu-nvidia.c b/drivers/iommu/arm-smmu-nvidia.c
> index d93ceda..a429b2c 100644
> --- a/drivers/iommu/arm-smmu-nvidia.c
> +++ b/drivers/iommu/arm-smmu-nvidia.c
> @@ -56,11 +56,43 @@ static void nsmmu_write_reg64(struct arm_smmu_device *smmu,
>   		writeq_relaxed(val, nsmmu_page(smmu, i, page) + offset);
>   }
>   
> +static void nsmmu_tlb_sync_wait(struct arm_smmu_device *smmu, int page,
> +				int sync, int status, int inst)
> +{
> +	u32 reg;
> +	unsigned int spin_cnt, delay;
> +
> +	for (delay = 1; delay < TLB_LOOP_TIMEOUT; delay *= 2) {
> +		for (spin_cnt = TLB_SPIN_COUNT; spin_cnt > 0; spin_cnt--) {
> +			reg = readl_relaxed(
> +			      nsmmu_page(smmu, inst, page) + status);
> +			if (!(reg & sTLBGSTATUS_GSACTIVE))
> +				return;
> +			cpu_relax();
> +		}
> +		udelay(delay);
> +	}
> +	dev_err_ratelimited(smmu->dev,
> +			    "TLB sync timed out -- SMMU may be deadlocked\n");
> +}
> +
> +static void nsmmu_tlb_sync(struct arm_smmu_device *smmu, int page,
> +			   int sync, int status)
> +{
> +	int i;
> +
> +	arm_smmu_writel(smmu, page, sync, 0);
> +
> +	for (i = 0; i < to_nsmmu(smmu)->num_inst; i++)

It might make more sense to make this the innermost loop, i.e.:

	for (i = 0; i < nsmmu->num_inst; i++)
		reg &= readl_relaxed(nsmmu_page(smmu, i, page)...

since polling the instances in parallel rather than in series seems like 
it might be a bit more efficient.

> +		nsmmu_tlb_sync_wait(smmu, page, sync, status, i);
> +}
> +
>   static const struct arm_smmu_impl nsmmu_impl = {
>   	.read_reg = nsmmu_read_reg,
>   	.write_reg = nsmmu_write_reg,
>   	.read_reg64 = nsmmu_read_reg64,
>   	.write_reg64 = nsmmu_write_reg64,
> +	.tlb_sync = nsmmu_tlb_sync,
>   };
>   
>   struct arm_smmu_device *nvidia_smmu_impl_init(struct arm_smmu_device *smmu)
> diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
> index 46e1641..f5454e71 100644
> --- a/drivers/iommu/arm-smmu.c
> +++ b/drivers/iommu/arm-smmu.c
> @@ -52,9 +52,6 @@
>    */
>   #define QCOM_DUMMY_VAL -1
>   
> -#define TLB_LOOP_TIMEOUT		1000000	/* 1s! */
> -#define TLB_SPIN_COUNT			10
> -
>   #define MSI_IOVA_BASE			0x8000000
>   #define MSI_IOVA_LENGTH			0x100000
>   
> @@ -244,6 +241,11 @@ static void __arm_smmu_tlb_sync(struct arm_smmu_device *smmu, int page,
>   	unsigned int spin_cnt, delay;
>   	u32 reg;
>   
> +	if (smmu->impl->tlb_sync) {
> +		smmu->impl->tlb_sync(smmu, page, sync, status);

What I'd hoped is that rather than needing a hook for this, you could 
just override smmu_domain->tlb_ops from .init_context to wire up the 
alternate .sync method directly. That would save this extra level of 
indirection.

Robin.

> +		return;
> +	}
> +
>   	arm_smmu_writel(smmu, page, sync, QCOM_DUMMY_VAL);
>   	for (delay = 1; delay < TLB_LOOP_TIMEOUT; delay *= 2) {
>   		for (spin_cnt = TLB_SPIN_COUNT; spin_cnt > 0; spin_cnt--) {
> diff --git a/drivers/iommu/arm-smmu.h b/drivers/iommu/arm-smmu.h
> index 9645bf1..d3217f1 100644
> --- a/drivers/iommu/arm-smmu.h
> +++ b/drivers/iommu/arm-smmu.h
> @@ -207,6 +207,8 @@ enum arm_smmu_cbar_type {
>   /* Maximum number of context banks per SMMU */
>   #define ARM_SMMU_MAX_CBS		128
>   
> +#define TLB_LOOP_TIMEOUT		1000000	/* 1s! */
> +#define TLB_SPIN_COUNT			10
>   
>   /* Shared driver definitions */
>   enum arm_smmu_arch_version {
> @@ -336,6 +338,8 @@ struct arm_smmu_impl {
>   	int (*cfg_probe)(struct arm_smmu_device *smmu);
>   	int (*reset)(struct arm_smmu_device *smmu);
>   	int (*init_context)(struct arm_smmu_domain *smmu_domain);
> +	void (*tlb_sync)(struct arm_smmu_device *smmu, int page, int sync,
> +			 int status);
>   };
>   
>   static inline void __iomem *arm_smmu_page(struct arm_smmu_device *smmu, int n)
>