Re: [PATCH v4 02/09] iommu/ipmmu-vmsa: Add optional root device feature

From: Robin Murphy
Date: Mon Jun 19 2017 - 13:19:29 EST

Next message: Simran Singhal: "[PATCH] staging: sm750fb: Remove typedefs from _logical_chip_type_t and _clock_type_t"
Previous message: Borislav Petkov: "Re: [PATCH v6 26/34] iommu/amd: Allow the AMD IOMMU to work with memory encryption"
In reply to: Magnus Damm: "[PATCH v4 02/09] iommu/ipmmu-vmsa: Add optional root device feature"
Next in thread: Magnus Damm: "[PATCH v4 03/09] iommu/ipmmu-vmsa: Enable multi context support"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On 19/06/17 10:14, Magnus Damm wrote:
> From: Magnus Damm <damm+renesas@xxxxxxxxxxxxx>
>
> Add root device handling to the IPMMU driver by allowing certain
> DT compat strings to enable has_cache_leaf_nodes that in turn will
> support both root devices with interrupts and leaf devices that
> face the actual IPMMU consumer devices.
>
> Signed-off-by: Magnus Damm <damm+renesas@xxxxxxxxxxxxx>
> ---
>
> Changes since V3:
> - Reworked root finding code to make it easier to follow, thanks Geert!
>
> Changes since V2:
> - Fixed a bug in ipmmu_find_root() when only leaf devices are present
> - Broke out __ipmmu_find_root() to allow ->xlate() check for root devices
>
> Changes since V1:
> - Moved patch to earlier in the series
> - Updated code to work with recent changes in:
> [PATCH v3 00/06] iommu/ipmmu-vmsa: IPMMU multi-arch update V3
>
> drivers/iommu/ipmmu-vmsa.c | 95 ++++++++++++++++++++++++++++++++++++--------
> 1 file changed, 78 insertions(+), 17 deletions(-)
>
> --- 0015/drivers/iommu/ipmmu-vmsa.c
> +++ work/drivers/iommu/ipmmu-vmsa.c 2017-06-19 13:59:41.050607110 +0900
> @@ -36,6 +36,7 @@
>
> struct ipmmu_features {
> bool use_ns_alias_offset;
> + bool has_cache_leaf_nodes;
> };
>
> struct ipmmu_vmsa_device {
> @@ -44,6 +45,7 @@ struct ipmmu_vmsa_device {
> struct iommu_device iommu;
> struct list_head list;
> const struct ipmmu_features *features;
> + bool is_leaf;
> unsigned int num_utlbs;
> spinlock_t lock; /* Protects ctx and domains[] */
> DECLARE_BITMAP(ctx, IPMMU_CTX_MAX);
> @@ -54,6 +56,7 @@ struct ipmmu_vmsa_device {
>
> struct ipmmu_vmsa_domain {
> struct ipmmu_vmsa_device *mmu;
> + struct ipmmu_vmsa_device *root;

Would it not make more sense for this to be a property of the
ipmmu_device itself, rather than per ipmmu_domain? I may of course have
got the wrong idea of the topology here, but it seems as if mmu->is_leaf
could be expressed as mmu->root == mmu vs. mmu->root == some_other_mmu,
at which point there's one less thing to worry about in the domain.

> struct iommu_domain io_domain;
>
> struct io_pgtable_cfg cfg;
> @@ -203,6 +206,44 @@ static struct ipmmu_vmsa_iommu_priv *to_
> #define IMUASID_ASID0_SHIFT 0
>
> /* -----------------------------------------------------------------------------
> + * Root device handling
> + */
> +
> +static bool ipmmu_is_root(struct ipmmu_vmsa_device *mmu)
> +{
> + if (mmu->features->has_cache_leaf_nodes)
> + return mmu->is_leaf ? false : true> + else
> + return true; /* older IPMMU hardware treated as single root */
> +}
> +
> +static struct ipmmu_vmsa_device *__ipmmu_find_root(void)
> +{
> + struct ipmmu_vmsa_device *mmu;
> + struct ipmmu_vmsa_device *root = NULL;
> +
> + spin_lock(&ipmmu_devices_lock);
> +
> + list_for_each_entry(mmu, &ipmmu_devices, list) {
> + if (ipmmu_is_root(mmu)) {
> + root = mmu;
> + break;
> + }
> + }
> +
> + spin_unlock(&ipmmu_devices_lock);
> + return root;
> +}

I wonder if it might be tidier to use driver_for_each_device() for this,
and remove the local list at the same time as its previous user.

Either way, what happens if things end up hapening in this order:

1: probe leaf IPMMU B
2: probe device X behind IPMMU B
3: create and attach default domain for device X
4: probe root IPMMU A

We know X will defer if B isn't ready, but it doesn't seem (at a glance,
admittedly) that there's anything to enforce the expected probe ordering
between A and B. This seems like another argument for moving the
root/leaf association up to the device level, such that B can look up A
once in its own probe routine, and defer itself if necessary.

> +
> +static struct ipmmu_vmsa_device *ipmmu_find_root(struct ipmmu_vmsa_device *leaf)
> +{
> + if (ipmmu_is_root(leaf))
> + return leaf;
> + else
> + return __ipmmu_find_root();

Well actually, looking at that, I think what we have here is a very
long-winded way of implementing this:

static ipmmu_vmsa_device *root;

Yuck.

> +}
> +
> +/* -----------------------------------------------------------------------------
> * Read/Write Access
> */
>
> @@ -219,13 +260,13 @@ static void ipmmu_write(struct ipmmu_vms
>
> static u32 ipmmu_ctx_read(struct ipmmu_vmsa_domain *domain, unsigned int reg)
> {
> - return ipmmu_read(domain->mmu, domain->context_id * IM_CTX_SIZE + reg);
> + return ipmmu_read(domain->root, domain->context_id * IM_CTX_SIZE + reg);
> }
>
> static void ipmmu_ctx_write(struct ipmmu_vmsa_domain *domain, unsigned int reg,
> u32 data)
> {
> - ipmmu_write(domain->mmu, domain->context_id * IM_CTX_SIZE + reg, data);
> + ipmmu_write(domain->root, domain->context_id * IM_CTX_SIZE + reg, data);
> }
>
> /* -----------------------------------------------------------------------------
> @@ -360,7 +401,7 @@ static int ipmmu_domain_init_context(str
> * TODO: Add support for coherent walk through CCI with DVM and remove
> * cache handling. For now, delegate it to the io-pgtable code.
> */
> - domain->cfg.iommu_dev = domain->mmu->dev;
> + domain->cfg.iommu_dev = domain->root->dev;
>
> domain->iop = alloc_io_pgtable_ops(ARM_32_LPAE_S1, &domain->cfg,
> domain);
> @@ -370,7 +411,7 @@ static int ipmmu_domain_init_context(str
> /*
> * Find an unused context.
> */
> - ret = ipmmu_domain_allocate_context(domain->mmu, domain);
> + ret = ipmmu_domain_allocate_context(domain->root, domain);
> if (ret == IPMMU_CTX_MAX) {
> free_io_pgtable_ops(domain->iop);
> return -EBUSY;
> @@ -441,7 +482,7 @@ static void ipmmu_domain_destroy_context
> */
> ipmmu_ctx_write(domain, IMCTR, IMCTR_FLUSH);
> ipmmu_tlb_sync(domain);
> - ipmmu_domain_free_context(domain->mmu, domain->context_id);
> + ipmmu_domain_free_context(domain->root, domain->context_id);
> }
>
> /* -----------------------------------------------------------------------------
> @@ -555,7 +596,7 @@ static int ipmmu_attach_device(struct io
> {
> struct ipmmu_vmsa_iommu_priv *priv = to_priv(dev);
> struct iommu_fwspec *fwspec = dev->iommu_fwspec;
> - struct ipmmu_vmsa_device *mmu = priv->mmu;
> + struct ipmmu_vmsa_device *root, *mmu = priv->mmu;
> struct ipmmu_vmsa_domain *domain = to_vmsa_domain(io_domain);
> unsigned long flags;
> unsigned int i;
> @@ -566,11 +607,18 @@ static int ipmmu_attach_device(struct io
> return -ENXIO;
> }
>
> + root = ipmmu_find_root(priv->mmu);
> + if (!root) {
> + dev_err(dev, "Unable to locate root IPMMU\n");
> + return -EAGAIN;
> + }
> +
> spin_lock_irqsave(&domain->lock, flags);
>
> if (!domain->mmu) {
> /* The domain hasn't been used yet, initialize it. */
> domain->mmu = mmu;
> + domain->root = root;
> ret = ipmmu_domain_init_context(domain);
> } else if (domain->mmu != mmu) {
> /*
> @@ -911,6 +959,7 @@ static void ipmmu_device_reset(struct ip
>
> static const struct ipmmu_features ipmmu_features_default = {
> .use_ns_alias_offset = true,
> + .has_cache_leaf_nodes = false,
> };
>
> static const struct of_device_id ipmmu_of_ids[] = {
> @@ -965,19 +1014,31 @@ static int ipmmu_probe(struct platform_d
> mmu->base += IM_NS_ALIAS_OFFSET;
>
> irq = platform_get_irq(pdev, 0);
> - if (irq < 0) {
> - dev_err(&pdev->dev, "no IRQ found\n");
> - return irq;
> - }
>
> - ret = devm_request_irq(&pdev->dev, irq, ipmmu_irq, 0,
> - dev_name(&pdev->dev), mmu);
> - if (ret < 0) {
> - dev_err(&pdev->dev, "failed to request IRQ %d\n", irq);
> - return ret;
> - }
> + /*
> + * Determine if this IPMMU instance is a leaf device by checking
> + * if the renesas,ipmmu-main property exists or not.
> + */
> + if (mmu->features->has_cache_leaf_nodes &&
> + of_find_property(pdev->dev.of_node, "renesas,ipmmu-main", NULL))
> + mmu->is_leaf = true;

Given my previous comments:

if (mmu->features->has_cache_leaf_nodes) {
root_np = of_parse_phandle(pdev->dev.of_node,
"renesas,ipmmu-main", 0)
if (root_np) {
root_pdev = of_find_device_by_phandle(root_np);
mmu->root = platform_get_drvdata(root_pdev);
} else {
mmu->root = mmu;
}
}

or something along those lines?

Robin.

> +
> + /* Root devices have mandatory IRQs */
> + if (ipmmu_is_root(mmu)) {
> + if (irq < 0) {
> + dev_err(&pdev->dev, "no IRQ found\n");
> + return irq;
> + }
>
> - ipmmu_device_reset(mmu);
> + ret = devm_request_irq(&pdev->dev, irq, ipmmu_irq, 0,
> + dev_name(&pdev->dev), mmu);
> + if (ret < 0) {
> + dev_err(&pdev->dev, "failed to request IRQ %d\n", irq);
> + return ret;
> + }
> +
> + ipmmu_device_reset(mmu);
> + }
>
> ret = iommu_device_register(&mmu->iommu);
> if (ret)
>

Next message: Simran Singhal: "[PATCH] staging: sm750fb: Remove typedefs from _logical_chip_type_t and _clock_type_t"
Previous message: Borislav Petkov: "Re: [PATCH v6 26/34] iommu/amd: Allow the AMD IOMMU to work with memory encryption"
In reply to: Magnus Damm: "[PATCH v4 02/09] iommu/ipmmu-vmsa: Add optional root device feature"
Next in thread: Magnus Damm: "[PATCH v4 03/09] iommu/ipmmu-vmsa: Enable multi context support"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]