arm: msm: Add MSM IOMMU support.

Add support for the IOMMUs found on the upcoming Qualcomm MSM8x60 chips. These IOMMUs allow virtualization of the address space used by most of the multimedia cores on these chips. The driver implements the kernel's IOMMU API.

arch/arm/mach-msm/iommu.c new file +104 -0

+			dmac_flush_range(sl_table, sl_table + SZ_4K);
+		}
+#endif
+
+	list_for_each_entry(ctx_drvdata, &priv->list_attached, attached_elm) {
+		if (!ctx_drvdata->pdev || !ctx_drvdata->pdev->dev.parent)
+			BUG();
+
+		iommu_drvdata = dev_get_drvdata(ctx_drvdata->pdev->dev.parent);
+		SET_CTX_TLBIALL(iommu_drvdata->base, ctx_drvdata->num, 0);
+	}
+}
+
+static void __reset_context(void __iomem *base, int ctx)
+{
+	SET_BPRCOSH(base, ctx, 0);
+	SET_BPRCISH(base, ctx, 0);
+	SET_BPRCNSH(base, ctx, 0);
+	SET_BPSHCFG(base, ctx, 0);
+	SET_BPMTCFG(base, ctx, 0);
+	SET_ACTLR(base, ctx, 0);
+	SET_SCTLR(base, ctx, 0);
+	SET_FSRRESTORE(base, ctx, 0);
+	SET_TTBR0(base, ctx, 0);
+	SET_TTBR1(base, ctx, 0);
+	SET_TTBCR(base, ctx, 0);
+	SET_BFBCR(base, ctx, 0);
+	SET_PAR(base, ctx, 0);
+	SET_FAR(base, ctx, 0);
+	SET_CTX_TLBIALL(base, ctx, 0);
+	SET_TLBFLPTER(base, ctx, 0);
+	SET_TLBSLPTER(base, ctx, 0);
+	SET_TLBLKCR(base, ctx, 0);
+	SET_PRRR(base, ctx, 0);
+	SET_NMRR(base, ctx, 0);
+	SET_CONTEXTIDR(base, ctx, 0);
+}
+
+static void __program_context(void __iomem *base, int ctx, phys_addr_t pgtable)
+{
+	__reset_context(base, ctx);
+
+	/* Set up HTW mode */
+	/* TLB miss configuration: perform HTW on miss */
+	SET_TLBMCFG(base, ctx, 0x3);
+
+	/* V2P configuration: HTW for access */
+	SET_V2PCFG(base, ctx, 0x3);
+
+	SET_TTBCR(base, ctx, 0);
+	SET_TTBR0_PA(base, ctx, (pgtable >> 14));
+
+	/* Invalidate the TLB for this context */
+	SET_CTX_TLBIALL(base, ctx, 0);
+
+	/* Set interrupt number to "secure" interrupt */
+	SET_IRPTNDX(base, ctx, 0);
+
+	/* Enable context fault interrupt */
+	SET_CFEIE(base, ctx, 1);
+
+	/* Stall access on a context fault and let the handler deal with it */
+	SET_CFCFG(base, ctx, 1);
+
+	/* Redirect all cacheable requests to L2 slave port. */
+	SET_RCISH(base, ctx, 1);
+	SET_RCOSH(base, ctx, 1);
+	SET_RCNSH(base, ctx, 1);
+
+	/* Turn on TEX Remap */
+	SET_TRE(base, ctx, 1);
+
+	/* Do not configure PRRR / NMRR on the IOMMU for now. We will assume
+	 * TEX class 0 for everything until attributes are properly worked out
+	 */
+	SET_PRRR(base, ctx, 0);
+	SET_NMRR(base, ctx, 0);
+
+	/* Turn on BFB prefetch */
+	SET_BFBDFE(base, ctx, 1);
+
+#ifdef CONFIG_IOMMU_PGTABLES_L2
+	/* Configure page tables as inner-cacheable and shareable to reduce
+	 * the TLB miss penalty.
+	 */
+	SET_TTBR0_SH(base, ctx, 1);
+	SET_TTBR1_SH(base, ctx, 1);
+
+	SET_TTBR0_NOS(base, ctx, 1);
+	SET_TTBR1_NOS(base, ctx, 1);
+
+	SET_TTBR0_IRGNH(base, ctx, 0); /* WB, WA */
+	SET_TTBR0_IRGNL(base, ctx, 1);
+
+	SET_TTBR1_IRGNH(base, ctx, 0); /* WB, WA */
+	SET_TTBR1_IRGNL(base, ctx, 1);
+
+	SET_TTBR0_ORGN(base, ctx, 1); /* WB, WA */
+	SET_TTBR1_ORGN(base, ctx, 1); /* WB, WA */
+#endif
+
+	/* Enable the MMU */
+	SET_M(base, ctx, 1);
+}

arch/arm/mach-msm/iommu.c new file +18 -0

+
+	if (!priv->pgtable)
+		goto fail_nomem;
+
+	memset(priv->pgtable, 0, SZ_16K);
+	domain->priv = priv;
+	return 0;
+
+fail_nomem:
+	kfree(priv);
+	return -ENOMEM;
+}
+
+static void msm_iommu_domain_destroy(struct iommu_domain *domain)
+{
+	struct msm_priv *priv;
+	unsigned long flags;
+	unsigned long *fl_table;

arch/arm/mach-msm/iommu.c new file +128 -0

+		priv->pgtable = NULL;
+	}
+
+	kfree(priv);
+	spin_unlock_irqrestore(&msm_iommu_lock, flags);
+}
+
+static int msm_iommu_attach_dev(struct iommu_domain *domain, struct device *dev)
+{
+	struct msm_priv *priv;
+	struct msm_iommu_ctx_dev *ctx_dev;
+	struct msm_iommu_drvdata *iommu_drvdata;
+	struct msm_iommu_ctx_drvdata *ctx_drvdata;
+	struct msm_iommu_ctx_drvdata *tmp_drvdata;
+	int ret = 0;
+	unsigned long flags;
+
+	spin_lock_irqsave(&msm_iommu_lock, flags);
+
+	priv = domain->priv;
+
+	if (!priv || !dev) {
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	iommu_drvdata = dev_get_drvdata(dev->parent);
+	ctx_drvdata = dev_get_drvdata(dev);
+	ctx_dev = dev->platform_data;
+
+	if (!iommu_drvdata || !ctx_drvdata || !ctx_dev) {
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	list_for_each_entry(tmp_drvdata, &priv->list_attached, attached_elm)
+		if (tmp_drvdata == ctx_drvdata) {
+			ret = -EBUSY;
+			goto fail;
+		}
+
+	__program_context(iommu_drvdata->base, ctx_dev->num,
+			  __pa(priv->pgtable));
+
+	list_add(&(ctx_drvdata->attached_elm), &priv->list_attached);
+	__flush_iotlb(domain);
+
+fail:
+	spin_unlock_irqrestore(&msm_iommu_lock, flags);
+	return ret;
+}
+
+static void msm_iommu_detach_dev(struct iommu_domain *domain,
+				 struct device *dev)
+{
+	struct msm_priv *priv;
+	struct msm_iommu_ctx_dev *ctx_dev;
+	struct msm_iommu_drvdata *iommu_drvdata;
+	struct msm_iommu_ctx_drvdata *ctx_drvdata;
+	unsigned long flags;
+
+	spin_lock_irqsave(&msm_iommu_lock, flags);
+	priv = domain->priv;
+
+	if (!priv || !dev)
+		goto fail;
+
+	iommu_drvdata = dev_get_drvdata(dev->parent);
+	ctx_drvdata = dev_get_drvdata(dev);
+	ctx_dev = dev->platform_data;
+
+	if (!iommu_drvdata || !ctx_drvdata || !ctx_dev)
+		goto fail;
+
+	__flush_iotlb(domain);
+	__reset_context(iommu_drvdata->base, ctx_dev->num);
+	list_del_init(&ctx_drvdata->attached_elm);
+
+fail:
+	spin_unlock_irqrestore(&msm_iommu_lock, flags);
+}
+
+static int msm_iommu_map(struct iommu_domain *domain, unsigned long va,
+			 phys_addr_t pa, int order, int prot)
+{
+	struct msm_priv *priv;
+	unsigned long flags;
+	unsigned long *fl_table;
+	unsigned long *fl_pte;
+	unsigned long fl_offset;
+	unsigned long *sl_table;
+	unsigned long *sl_pte;
+	unsigned long sl_offset;
+	size_t len = 0x1000UL << order;
+	int ret = 0;
+
+	spin_lock_irqsave(&msm_iommu_lock, flags);
+	priv = domain->priv;
+
+	if (!priv) {
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	fl_table = priv->pgtable;
+
+	if (len != SZ_16M && len != SZ_1M &&
+	    len != SZ_64K && len != SZ_4K) {
+		pr_debug("Bad size: %d\n", len);
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	if (!fl_table) {
+		pr_debug("Null page table\n");
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	fl_offset = FL_OFFSET(va);	/* Upper 12 bits */
+	fl_pte = fl_table + fl_offset;	/* int pointers, 4 bytes */
+
+	if (len == SZ_16M) {
+		int i = 0;
+		for (i = 0; i < 16; i++)
+			*(fl_pte+i) = (pa & 0xFF000000) | FL_SUPERSECTION |
+				  FL_AP_READ | FL_AP_WRITE | FL_TYPE_SECT |
+				  FL_SHARED;

arch/arm/mach-msm/iommu.c new file +137 -0

+
+		if (!sl) {
+			pr_debug("Could not allocate second level table\n");
+			ret = -ENOMEM;
+			goto fail;
+		}
+
+		memset(sl, 0, SZ_4K);
+		*fl_pte = ((((int)__pa(sl)) & 0xFFFFFC00) | FL_TYPE_TABLE);
+	}
+
+	sl_table = (unsigned long *) __va(((*fl_pte) & 0xFFFFFC00));
+	sl_offset = SL_OFFSET(va);
+	sl_pte = sl_table + sl_offset;
+
+
+	if (len == SZ_4K)
+		*sl_pte = (pa & 0xFFFFF000) | SL_AP0 | SL_AP1 |	SL_SHARED |
+					      SL_TYPE_SMALL;
+
+	if (len == SZ_64K) {
+		int i;
+
+		for (i = 0; i < 16; i++)
+			*(sl_pte+i) = (pa & 0xFFFF0000) | SL_AP0 | SL_AP1 |
+					      SL_SHARED | SL_TYPE_LARGE;
+	}
+
+	__flush_iotlb(domain);
+fail:
+	spin_unlock_irqrestore(&msm_iommu_lock, flags);
+	return ret;
+}
+
+static int msm_iommu_unmap(struct iommu_domain *domain, unsigned long va,
+			    int order)
+{
+	struct msm_priv *priv;
+	unsigned long flags;
+	unsigned long *fl_table;
+	unsigned long *fl_pte;
+	unsigned long fl_offset;
+	unsigned long *sl_table;
+	unsigned long *sl_pte;
+	unsigned long sl_offset;
+	size_t len = 0x1000UL << order;
+	int i, ret = 0;
+
+	spin_lock_irqsave(&msm_iommu_lock, flags);
+
+	priv = domain->priv;
+
+	if (!priv) {
+		ret = -ENODEV;
+		goto fail;
+	}
+
+	fl_table = priv->pgtable;
+
+	if (len != SZ_16M && len != SZ_1M &&
+	    len != SZ_64K && len != SZ_4K) {
+		pr_debug("Bad length: %d\n", len);
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	if (!fl_table) {
+		pr_debug("Null page table\n");
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	fl_offset = FL_OFFSET(va);	/* Upper 12 bits */
+	fl_pte = fl_table + fl_offset;	/* int pointers, 4 bytes */
+
+	if (*fl_pte == 0) {
+		pr_debug("First level PTE is 0\n");
+		ret = -ENODEV;
+		goto fail;
+	}
+
+	/* Unmap supersection */
+	if (len == SZ_16M)
+		for (i = 0; i < 16; i++)
+			*(fl_pte+i) = 0;
+
+	if (len == SZ_1M)
+		*fl_pte = 0;
+
+	sl_table = (unsigned long *) __va(((*fl_pte) & 0xFFFFFC00));
+	sl_offset = SL_OFFSET(va);
+	sl_pte = sl_table + sl_offset;
+
+	if (len == SZ_64K) {
+		for (i = 0; i < 16; i++)
+			*(sl_pte+i) = 0;
+	}
+
+	if (len == SZ_4K)
+		*sl_pte = 0;
+
+	if (len == SZ_4K || len == SZ_64K) {
+		int used = 0;
+
+		for (i = 0; i < 256; i++)
+			if (sl_table[i])
+				used = 1;
+		if (!used) {
+			free_page((unsigned long)sl_table);
+			*fl_pte = 0;
+		}
+	}
+
+	__flush_iotlb(domain);
+fail:
+	spin_unlock_irqrestore(&msm_iommu_lock, flags);
+	return ret;
+}
+
+static phys_addr_t msm_iommu_iova_to_phys(struct iommu_domain *domain,
+					  unsigned long va)
+{
+	struct msm_priv *priv;
+	struct msm_iommu_drvdata *iommu_drvdata;
+	struct msm_iommu_ctx_drvdata *ctx_drvdata;
+	unsigned int par;
+	unsigned long flags;
+	void __iomem *base;
+	phys_addr_t ret = 0;
+	int ctx;
+
+	spin_lock_irqsave(&msm_iommu_lock, flags);
+
+	priv = domain->priv;
+	if (list_empty(&priv->list_attached))
+		goto fail;
+

I did a very brief look.

arch/arm/mach-msm/iommu.c

Quoted 3 lines

arch/arm/mach-msm/iommu.c new file +3 -0

+	dmac_flush_range(fl_table, fl_table + SZ_16K);
+
+	for (i = 0; i<  4096; i++)

Can you use a #define for the 4096? Will that value ever change?

Review

Quoted 2 lines

(quoted diff) +2 -0

+		if ((fl_table[i]&  0x03) == FL_TYPE_TABLE) {
+			void *sl_table = __va(fl_table[i]&  0xFFFFFC00);

Does it make sense to define say:
#define MASK_XX (x) x& 0xFFFFFC00
and
#define MASK_XY (x) x& 0xFFFF0000

and use those instead of these hard-coded values?

arch/arm/mach-msm/iommu.c line 174

Quoted 1 line

arch/arm/mach-msm/iommu.c new file +1 -0

+	priv->pgtable = (unsigned long *)__get_free_pages(GFP_KERNEL, 2);

Ugh. Can you use get_order(SZ_16K) instead?

arch/arm/mach-msm/iommu.c

Quoted 3 lines

arch/arm/mach-msm/iommu.c new file +3 -0

+	/* Invalidate context TLB */
+	SET_CTX_TLBIALL(base, ctx, 0);
+	SET_V2PPR_VA(base, ctx, va>>  12);

Hmm, PAGE_SHIFT can't be used? Or some other #define?