[RFC v01 0/3] xen/arm: introduce IOMMU driver for OMAP platforms

[RFC v01 0/3] xen/arm: introduce IOMMU driver for OMAP platforms

[Andrii Tseglytskyi]

Hi,

The following patch series is an RFC for possible implementation of simple MMU module,
which is designed to translate IPA to MA for peripheral processors like GPU / IPU
for OMAP platforms. Currently on our OMAP platform (OMAP5 panda) we have 3 external MMUs
which need to be handled properly.

It would be great to get a community feedback - will this be useful for Xen project?

Let me describe an algorithm briefly. It is simple and straightforward.
The following simple logic is used to translate addresses from IPA to MA:

1. During boot time guest domain creates "pagetable" for external MMU IP.
Pagetable is a singletone data structure, which is stored in ususal kernel
heap memory. All memory mappings for corresponding MMU are stored inside it.
Format of "pagetable" is well defined.

2. Guest domain enables peripheral remote processor. As a part of enable sequence
kernel allocates chunks of heap memory needed for remote processor and stores
pointers to allocated chunks in already created "pagetable". After it writes
a physical address of pagetable to MMU configuration register. As result MMU IP
knows about all allocations, and remote processor can use them directly in its
software.

3. Xen omap mmu driver creates a trap for access to MMU configuration registers.
It reads a physical address of "pagetable" from MMU register and creates a copy
of it in own memory. As result - we have two similar configuration data structures -
first - in guest domain kernel, second - in Xen hypervisor.

4. Xen omap mmu driver parses its own copy of pagetable and translate all physical
addresses to corresponding machine addresses using existing p2m API call.
After it writes a physical address  of its pagetable (with already translated PA to MA)
to MMU IP configuration registers and returns control to guest domain.

As a result - guest domain continues enabling remote processor with it MMU and MMU
will use new pagetable, modified by Xen omap mmu driver. New pagetable will be used
directly by MMU IP, and its new structure will be hidden for guest domain kernel,
it won't know anything about p2m translation.

Verified with Xen 4.4-unstable, Linux kernel 3.8 as Dom0, Linux(Android) kernel 3.4 as DomU.
Target platform OMAP5 panda.

Thank you for your attention,

Regards,

Andrii Tseglytskyi (3):
  arm: omap: introduce iommu module
  arm: omap: translate iommu mapping to 4K pages
  arm: omap: cleanup iopte allocations

 xen/arch/arm/Makefile     |    1 +
 xen/arch/arm/io.c         |    1 +
 xen/arch/arm/io.h         |    1 +
 xen/arch/arm/omap_iommu.c |  492 +++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 495 insertions(+)
 create mode 100644 xen/arch/arm/omap_iommu.c

-- 
1.7.9.5

[RFC v01 1/3] arm: omap: introduce iommu module

[Andrii Tseglytskyi]

omap IOMMU module is designed to handle access to external
omap MMUs, connected to the L3 bus.

Change-Id: I96bbf2738e9dd2e21662e0986ca15c60183e669e
Signed-off-by: Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
---
 xen/arch/arm/Makefile     |    1 +
 xen/arch/arm/io.c         |    1 +
 xen/arch/arm/io.h         |    1 +
 xen/arch/arm/omap_iommu.c |  415 +++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 418 insertions(+)
 create mode 100644 xen/arch/arm/omap_iommu.c

diff --git a/xen/arch/arm/Makefile b/xen/arch/arm/Makefile
index 003ac84..cb0b385 100644
--- a/xen/arch/arm/Makefile
+++ b/xen/arch/arm/Makefile
@@ -14,6 +14,7 @@ obj-y += io.o
 obj-y += irq.o
 obj-y += kernel.o
 obj-y += mm.o
+obj-y += omap_iommu.o
 obj-y += p2m.o
 obj-y += percpu.o
 obj-y += guestcopy.o
diff --git a/xen/arch/arm/io.c b/xen/arch/arm/io.c
index a6db00b..3281b67 100644
--- a/xen/arch/arm/io.c
+++ b/xen/arch/arm/io.c
@@ -26,6 +26,7 @@ static const struct mmio_handler *const mmio_handlers[] =
 {
     &vgic_distr_mmio_handler,
     &vuart_mmio_handler,
+	&mmu_mmio_handler,
 };
 #define MMIO_HANDLER_NR ARRAY_SIZE(mmio_handlers)
 
diff --git a/xen/arch/arm/io.h b/xen/arch/arm/io.h
index 8d252c0..acb5dff 100644
--- a/xen/arch/arm/io.h
+++ b/xen/arch/arm/io.h
@@ -42,6 +42,7 @@ struct mmio_handler {
 
 extern const struct mmio_handler vgic_distr_mmio_handler;
 extern const struct mmio_handler vuart_mmio_handler;
+extern const struct mmio_handler mmu_mmio_handler;
 
 extern int handle_mmio(mmio_info_t *info);
 
diff --git a/xen/arch/arm/omap_iommu.c b/xen/arch/arm/omap_iommu.c
new file mode 100644
index 0000000..4dab30f
--- /dev/null
+++ b/xen/arch/arm/omap_iommu.c
@@ -0,0 +1,415 @@
+/*
+ * xen/arch/arm/omap_iommu.c
+ *
+ * Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
+ * Copyright (c) 2013 GlobalLogic
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#include <xen/config.h>
+#include <xen/lib.h>
+#include <xen/errno.h>
+#include <xen/mm.h>
+#include <xen/vmap.h>
+#include <xen/init.h>
+#include <xen/sched.h>
+#include <xen/stdbool.h>
+#include <asm/system.h>
+#include <asm/current.h>
+#include <asm/io.h>
+#include <asm/p2m.h>
+
+#include "io.h"
+
+/* register where address of page table is stored */
+#define MMU_TTB			0x4c
+
+/*
+ * "L2 table" address mask and size definitions.
+ */
+
+/* 1st level translation */
+#define IOPGD_SHIFT		20
+#define IOPGD_SIZE		(1UL << IOPGD_SHIFT)
+#define IOPGD_MASK		(~(IOPGD_SIZE - 1))
+
+/* "supersection" - 16 Mb */
+#define IOSUPER_SHIFT		24
+#define IOSUPER_SIZE		(1UL << IOSUPER_SHIFT)
+#define IOSUPER_MASK		(~(IOSUPER_SIZE - 1))
+
+/* "section"  - 1 Mb */
+#define IOSECTION_SHIFT		20
+#define IOSECTION_SIZE		(1UL << IOSECTION_SHIFT)
+#define IOSECTION_MASK		(~(IOSECTION_SIZE - 1))
+
+/* 4096 first level descriptors for "supersection" and "section" */
+#define PTRS_PER_IOPGD		(1UL << (32 - IOPGD_SHIFT))
+#define IOPGD_TABLE_SIZE	(PTRS_PER_IOPGD * sizeof(u32))
+
+/* 2nd level translation */
+
+/* "small page" - 4Kb */
+#define IOPTE_SMALL_SHIFT		12
+#define IOPTE_SMALL_SIZE		(1UL << IOPTE_SMALL_SHIFT)
+#define IOPTE_SMALL_MASK		(~(IOPTE_SMALL_SIZE - 1))
+
+/* "large page" - 64 Kb */
+#define IOPTE_LARGE_SHIFT		16
+#define IOPTE_LARGE_SIZE		(1UL << IOPTE_LARGE_SHIFT)
+#define IOPTE_LARGE_MASK		(~(IOPTE_LARGE_SIZE - 1))
+
+/* 256 second level descriptors for "small" and "large" pages */
+#define PTRS_PER_IOPTE		(1UL << (IOPGD_SHIFT - IOPTE_SMALL_SHIFT))
+#define IOPTE_TABLE_SIZE	(PTRS_PER_IOPTE * sizeof(u32))
+
+/*
+ * some descriptor attributes.
+ */
+#define IOPGD_TABLE		(1 << 0)
+#define IOPGD_SECTION	(2 << 0)
+#define IOPGD_SUPER		(1 << 18 | 2 << 0)
+
+#define iopgd_is_table(x)	(((x) & 3) == IOPGD_TABLE)
+#define iopgd_is_section(x)	(((x) & (1 << 18 | 3)) == IOPGD_SECTION)
+#define iopgd_is_super(x)	(((x) & (1 << 18 | 3)) == IOPGD_SUPER)
+
+#define IOPTE_SMALL		(2 << 0)
+#define IOPTE_LARGE		(1 << 0)
+
+#define iopte_is_small(x)	(((x) & 2) == IOPTE_SMALL)
+#define iopte_is_large(x)	(((x) & 3) == IOPTE_LARGE)
+#define iopte_offset(x)		((x) & IOPTE_SMALL_MASK)
+
+struct mmu_info {
+	const char			*name;
+	paddr_t				mem_start;
+	u32					mem_size;
+	u32					*pagetable;
+	void __iomem		*mem_map;
+};
+
+static struct mmu_info omap_ipu_mmu = {
+	.name		= "IPU_L2_MMU",
+	.mem_start	= 0x55082000,
+	.mem_size	= 0x1000,
+	.pagetable	= NULL,
+};
+
+static struct mmu_info omap_dsp_mmu = {
+	.name		= "DSP_L2_MMU",
+	.mem_start	= 0x4a066000,
+	.mem_size	= 0x1000,
+	.pagetable	= NULL,
+};
+
+static struct mmu_info *mmu_list[] = {
+	&omap_ipu_mmu,
+	&omap_dsp_mmu,
+};
+
+#define mmu_for_each(pfunc, data)						\
+({														\
+	u32 __i;											\
+	int __res = 0;										\
+														\
+	for (__i = 0; __i < ARRAY_SIZE(mmu_list); __i++) {	\
+		__res |= pfunc(mmu_list[__i], data);			\
+	}													\
+	__res;												\
+})
+
+static int mmu_check_mem_range(struct mmu_info *mmu, paddr_t addr)
+{
+	if ((addr >= mmu->mem_start) && (addr < (mmu->mem_start + mmu->mem_size)))
+		return 1;
+
+	return 0;
+}
+
+static inline struct mmu_info *mmu_lookup(u32 addr)
+{
+	u32 i;
+
+	for (i = 0; i < ARRAY_SIZE(mmu_list); i++) {
+		if (mmu_check_mem_range(mmu_list[i], addr))
+			return mmu_list[i];
+	}
+
+	return NULL;
+}
+
+static inline u32 mmu_virt_to_phys(u32 reg, u32 va, u32 mask)
+{
+	return (reg & mask) | (va & (~mask));
+}
+
+static inline u32 mmu_phys_to_virt(u32 reg, u32 pa, u32 mask)
+{
+	return (reg & ~mask) | pa;
+}
+
+static int mmu_mmio_check(struct vcpu *v, paddr_t addr)
+{
+	return mmu_for_each(mmu_check_mem_range, addr);
+}
+
+static int mmu_copy_pagetable(struct mmu_info *mmu)
+{
+	void __iomem *pagetable = NULL;
+	u32 pgaddr;
+
+	ASSERT(mmu);
+
+	/* read address where kernel MMU pagetable is stored */
+	pgaddr = readl(mmu->mem_map + MMU_TTB);
+	pagetable = ioremap(pgaddr, IOPGD_TABLE_SIZE);
+	if (!pagetable) {
+		printk("%s: %s failed to map pagetable\n",
+			   __func__, mmu->name);
+		return -EINVAL;
+	}
+
+	/*
+	 * pagetable can be changed since last time
+	 * we accessed it therefore we need to copy it each time
+	 */
+	memcpy(mmu->pagetable, pagetable, IOPGD_TABLE_SIZE);
+
+	iounmap(pagetable);
+
+	return 0;
+}
+
+#define mmu_dump_pdentry(da, iopgd, paddr, maddr, vaddr, mask)									\
+{																								\
+	const char *sect_type = (iopgd_is_table(iopgd) || (mask == IOPTE_SMALL_MASK) ||				\
+							(mask == IOPTE_LARGE_MASK)) ? "table"								\
+							: iopgd_is_super(iopgd) ? "supersection"							\
+							: iopgd_is_section(iopgd) ? "section"								\
+							: "Unknown section";												\
+	printk("[iopgd] %s da 0x%08x iopgd 0x%08x paddr 0x%08x maddr 0x%pS vaddr 0x%08x mask 0x%08x\n",\
+		   sect_type, da, iopgd, paddr, _p(maddr), vaddr, mask);								\
+}
+
+static u32 mmu_translate_pgentry(struct domain *dom, u32 iopgd, u32 da, u32 mask)
+{
+	u32 vaddr, paddr;
+	paddr_t maddr;
+
+	paddr = mmu_virt_to_phys(iopgd, da, mask);
+	maddr = p2m_lookup(dom, paddr);
+	vaddr = mmu_phys_to_virt(iopgd, maddr, mask);
+
+	return vaddr;
+}
+
+/*
+ * on boot table is empty
+ */
+static int mmu_translate_pagetable(struct domain *dom, struct mmu_info *mmu)
+{
+	u32 i;
+	int res;
+	bool table_updated = false;
+
+	ASSERT(dom);
+	ASSERT(mmu);
+
+	/* copy pagetable from  domain to xen */
+	res = mmu_copy_pagetable(mmu);
+	if (res) {
+		printk("%s: %s failed to map pagetable memory\n",
+			   __func__, mmu->name);
+		return res;
+	}
+
+	/* 1-st level translation */
+	for (i = 0; i < PTRS_PER_IOPGD; i++) {
+		u32 da;
+		u32 iopgd = mmu->pagetable[i];
+
+		if (!iopgd)
+			continue;
+
+		table_updated = true;
+
+		/* "supersection" 16 Mb */
+		if (iopgd_is_super(iopgd)) {
+			da = i << IOSECTION_SHIFT;
+			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSUPER_MASK);
+
+		/* "section" 1Mb */
+		} else if (iopgd_is_section(iopgd)) {
+			da = i << IOSECTION_SHIFT;
+			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
+
+		/* "table" */
+		} else if (iopgd_is_table(iopgd)) {
+			u32 j, mask;
+			u32 iopte = iopte_offset(iopgd);
+
+			/* 2-nd level translation */
+			for (j = 0; j < PTRS_PER_IOPTE; j++, iopte += IOPTE_SMALL_SIZE) {
+
+				/* "small table" 4Kb */
+				if (iopte_is_small(iopgd)) {
+					da = (i << IOSECTION_SHIFT) + (j << IOPTE_SMALL_SHIFT);
+					mask = IOPTE_SMALL_MASK;
+
+				/* "large table" 64Kb */
+				} else if (iopte_is_large(iopgd)) {
+					da = (i << IOSECTION_SHIFT) + (j << IOPTE_LARGE_SHIFT);
+					mask = IOPTE_LARGE_MASK;
+
+				/* error */
+				} else {
+					printk("%s Unknown table type 0x%08x\n", mmu->name, iopte);
+					return -EINVAL;
+				}
+
+				/* translate 2-nd level entry */
+				mmu->pagetable[i] = mmu_translate_pgentry(dom, iopte, da, mask);
+			}
+
+			continue;
+
+		/* error */
+		} else {
+			printk("%s Unknown entry 0x%08x\n", mmu->name, iopgd);
+			return -EINVAL;
+		}
+	}
+
+	/* force omap IOMMU to use new pagetable */
+	if (table_updated) {
+		paddr_t maddr;
+		flush_xen_dcache_va_range(mmu->pagetable, IOPGD_TABLE_SIZE);
+		maddr = __pa(mmu->pagetable);
+		writel(maddr, mmu->mem_map + MMU_TTB);
+		printk("%s update pagetable, maddr 0x%pS\n", mmu->name, _p(maddr));
+	}
+
+	return 0;
+}
+
+static int mmu_trap_write_access(struct domain *dom,
+								 struct mmu_info *mmu, mmio_info_t *info)
+{
+	struct cpu_user_regs *regs = guest_cpu_user_regs();
+	register_t *r = select_user_reg(regs, info->dabt.reg);
+	int res = 0;
+
+	switch (info->gpa - mmu->mem_start) {
+		case MMU_TTB:
+			printk("%s MMU_TTB write access 0x%pS <= 0x%08x\n",
+				   mmu->name, _p(info->gpa), *r);
+			res = mmu_translate_pagetable(dom, mmu);
+			break;
+		default:
+			break;
+	}
+
+	return res;
+}
+
+static int mmu_mmio_read(struct vcpu *v, mmio_info_t *info)
+{
+	struct mmu_info *mmu = NULL;
+    struct cpu_user_regs *regs = guest_cpu_user_regs();
+    register_t *r = select_user_reg(regs, info->dabt.reg);
+
+	mmu = mmu_lookup(info->gpa);
+	if (!mmu) {
+		printk("%s: can't get mmu for addr 0x%08x\n", __func__, (u32)info->gpa);
+		return -EINVAL;
+	}
+
+    *r = readl(mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
+
+    return 1;
+}
+
+static int mmu_mmio_write(struct vcpu *v, mmio_info_t *info)
+{
+	struct domain *dom = v->domain;
+	struct mmu_info *mmu = NULL;
+    struct cpu_user_regs *regs = guest_cpu_user_regs();
+    register_t *r = select_user_reg(regs, info->dabt.reg);
+	int res;
+
+	mmu = mmu_lookup(info->gpa);
+	if (!mmu) {
+		printk("%s: can't get mmu for addr 0x%08x\n", __func__, (u32)info->gpa);
+		return -EINVAL;
+	}
+
+	/*
+	 * make sure that user register is written first in this function
+	 * following calls may expect valid data in it
+	 */
+    writel(*r, mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
+
+	res = mmu_trap_write_access(dom, mmu, info);
+	if (res)
+		return res;
+
+    return 1;
+}
+
+static int mmu_init(struct mmu_info *mmu, u32 data)
+{
+	ASSERT(mmu);
+	ASSERT(!mmu->mem_map);
+	ASSERT(!mmu->pagetable);
+
+    mmu->mem_map = ioremap(mmu->mem_start, mmu->mem_size);
+	if (!mmu->mem_map) {
+		printk("%s: %s failed to map memory\n",  __func__, mmu->name);
+		return -EINVAL;
+	}
+
+	printk("%s: %s ipu_map = 0x%pS\n", __func__, mmu->name, _p(mmu->mem_map));
+
+	mmu->pagetable = xzalloc_bytes(IOPGD_TABLE_SIZE);
+	if (!mmu->pagetable) {
+		printk("%s: %s failed to alloc private pagetable\n",
+			   __func__, mmu->name);
+		return -ENOMEM;
+	}
+
+	printk("%s: %s private pagetable %lu bytes\n",
+		   __func__, mmu->name, IOPGD_TABLE_SIZE);
+
+	return 0;
+}
+
+static int mmu_init_all(void)
+{
+	int res;
+
+	res = mmu_for_each(mmu_init, 0);
+	if (res) {
+		printk("%s error during init %d\n", __func__, res);
+		return res;
+	}
+
+	return 0;
+}
+
+const struct mmio_handler mmu_mmio_handler = {
+	.check_handler = mmu_mmio_check,
+	.read_handler  = mmu_mmio_read,
+	.write_handler = mmu_mmio_write,
+};
+
+__initcall(mmu_init_all);
-- 
1.7.9.5

Re: [RFC v01 1/3] arm: omap: introduce iommu module

[Stefano Stabellini]

On Wed, 22 Jan 2014, Andrii Tseglytskyi wrote:
> omap IOMMU module is designed to handle access to external
> omap MMUs, connected to the L3 bus.
> 
> Change-Id: I96bbf2738e9dd2e21662e0986ca15c60183e669e
> Signed-off-by: Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
> ---
>  xen/arch/arm/Makefile     |    1 +
>  xen/arch/arm/io.c         |    1 +
>  xen/arch/arm/io.h         |    1 +
>  xen/arch/arm/omap_iommu.c |  415 +++++++++++++++++++++++++++++++++++++++++++++
>  4 files changed, 418 insertions(+)
>  create mode 100644 xen/arch/arm/omap_iommu.c
> 
> diff --git a/xen/arch/arm/Makefile b/xen/arch/arm/Makefile
> index 003ac84..cb0b385 100644
> --- a/xen/arch/arm/Makefile
> +++ b/xen/arch/arm/Makefile
> @@ -14,6 +14,7 @@ obj-y += io.o
>  obj-y += irq.o
>  obj-y += kernel.o
>  obj-y += mm.o
> +obj-y += omap_iommu.o
>  obj-y += p2m.o
>  obj-y += percpu.o
>  obj-y += guestcopy.o
> diff --git a/xen/arch/arm/io.c b/xen/arch/arm/io.c
> index a6db00b..3281b67 100644
> --- a/xen/arch/arm/io.c
> +++ b/xen/arch/arm/io.c
> @@ -26,6 +26,7 @@ static const struct mmio_handler *const mmio_handlers[] =
>  {
>      &vgic_distr_mmio_handler,
>      &vuart_mmio_handler,
> +	&mmu_mmio_handler,
>  };
>  #define MMIO_HANDLER_NR ARRAY_SIZE(mmio_handlers)

I think that omap_iommu should be a platform specific driver, and it
should hook into a set of platform specific mmio_handlers instead of
using the generic mmio_handler structure.


> diff --git a/xen/arch/arm/io.h b/xen/arch/arm/io.h
> index 8d252c0..acb5dff 100644
> --- a/xen/arch/arm/io.h
> +++ b/xen/arch/arm/io.h
> @@ -42,6 +42,7 @@ struct mmio_handler {
>  
>  extern const struct mmio_handler vgic_distr_mmio_handler;
>  extern const struct mmio_handler vuart_mmio_handler;
> +extern const struct mmio_handler mmu_mmio_handler;
>  
>  extern int handle_mmio(mmio_info_t *info);
>  
> diff --git a/xen/arch/arm/omap_iommu.c b/xen/arch/arm/omap_iommu.c
> new file mode 100644
> index 0000000..4dab30f
> --- /dev/null
> +++ b/xen/arch/arm/omap_iommu.c

It should probably live under xen/arch/arm/platforms.


> @@ -0,0 +1,415 @@
> +/*
> + * xen/arch/arm/omap_iommu.c
> + *
> + * Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
> + * Copyright (c) 2013 GlobalLogic
> + *
> + * This program is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License as published by
> + * the Free Software Foundation; either version 2 of the License, or
> + * (at your option) any later version.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> + * GNU General Public License for more details.
> + */
> +
> +#include <xen/config.h>
> +#include <xen/lib.h>
> +#include <xen/errno.h>
> +#include <xen/mm.h>
> +#include <xen/vmap.h>
> +#include <xen/init.h>
> +#include <xen/sched.h>
> +#include <xen/stdbool.h>
> +#include <asm/system.h>
> +#include <asm/current.h>
> +#include <asm/io.h>
> +#include <asm/p2m.h>
> +
> +#include "io.h"
> +
> +/* register where address of page table is stored */
> +#define MMU_TTB			0x4c
> +
> +/*
> + * "L2 table" address mask and size definitions.
> + */
> +
> +/* 1st level translation */
> +#define IOPGD_SHIFT		20
> +#define IOPGD_SIZE		(1UL << IOPGD_SHIFT)
> +#define IOPGD_MASK		(~(IOPGD_SIZE - 1))
> +
> +/* "supersection" - 16 Mb */
> +#define IOSUPER_SHIFT		24
> +#define IOSUPER_SIZE		(1UL << IOSUPER_SHIFT)
> +#define IOSUPER_MASK		(~(IOSUPER_SIZE - 1))
> +
> +/* "section"  - 1 Mb */
> +#define IOSECTION_SHIFT		20
> +#define IOSECTION_SIZE		(1UL << IOSECTION_SHIFT)
> +#define IOSECTION_MASK		(~(IOSECTION_SIZE - 1))
> +
> +/* 4096 first level descriptors for "supersection" and "section" */
> +#define PTRS_PER_IOPGD		(1UL << (32 - IOPGD_SHIFT))
> +#define IOPGD_TABLE_SIZE	(PTRS_PER_IOPGD * sizeof(u32))
> +
> +/* 2nd level translation */
> +
> +/* "small page" - 4Kb */
> +#define IOPTE_SMALL_SHIFT		12
> +#define IOPTE_SMALL_SIZE		(1UL << IOPTE_SMALL_SHIFT)
> +#define IOPTE_SMALL_MASK		(~(IOPTE_SMALL_SIZE - 1))
> +
> +/* "large page" - 64 Kb */
> +#define IOPTE_LARGE_SHIFT		16
> +#define IOPTE_LARGE_SIZE		(1UL << IOPTE_LARGE_SHIFT)
> +#define IOPTE_LARGE_MASK		(~(IOPTE_LARGE_SIZE - 1))
> +
> +/* 256 second level descriptors for "small" and "large" pages */
> +#define PTRS_PER_IOPTE		(1UL << (IOPGD_SHIFT - IOPTE_SMALL_SHIFT))
> +#define IOPTE_TABLE_SIZE	(PTRS_PER_IOPTE * sizeof(u32))
> +
> +/*
> + * some descriptor attributes.
> + */
> +#define IOPGD_TABLE		(1 << 0)
> +#define IOPGD_SECTION	(2 << 0)
> +#define IOPGD_SUPER		(1 << 18 | 2 << 0)
> +
> +#define iopgd_is_table(x)	(((x) & 3) == IOPGD_TABLE)
> +#define iopgd_is_section(x)	(((x) & (1 << 18 | 3)) == IOPGD_SECTION)
> +#define iopgd_is_super(x)	(((x) & (1 << 18 | 3)) == IOPGD_SUPER)
> +
> +#define IOPTE_SMALL		(2 << 0)
> +#define IOPTE_LARGE		(1 << 0)
> +
> +#define iopte_is_small(x)	(((x) & 2) == IOPTE_SMALL)
> +#define iopte_is_large(x)	(((x) & 3) == IOPTE_LARGE)
> +#define iopte_offset(x)		((x) & IOPTE_SMALL_MASK)
> +
> +struct mmu_info {
> +	const char			*name;
> +	paddr_t				mem_start;
> +	u32					mem_size;
> +	u32					*pagetable;
> +	void __iomem		*mem_map;
> +};
> +
> +static struct mmu_info omap_ipu_mmu = {
> +	.name		= "IPU_L2_MMU",
> +	.mem_start	= 0x55082000,
> +	.mem_size	= 0x1000,
> +	.pagetable	= NULL,
> +};
> +
> +static struct mmu_info omap_dsp_mmu = {
> +	.name		= "DSP_L2_MMU",
> +	.mem_start	= 0x4a066000,
> +	.mem_size	= 0x1000,
> +	.pagetable	= NULL,
> +};
> +
> +static struct mmu_info *mmu_list[] = {
> +	&omap_ipu_mmu,
> +	&omap_dsp_mmu,
> +};
> +
> +#define mmu_for_each(pfunc, data)						\
> +({														\
> +	u32 __i;											\
> +	int __res = 0;										\
> +														\
> +	for (__i = 0; __i < ARRAY_SIZE(mmu_list); __i++) {	\
> +		__res |= pfunc(mmu_list[__i], data);			\
> +	}													\
> +	__res;												\
> +})
> +
> +static int mmu_check_mem_range(struct mmu_info *mmu, paddr_t addr)
> +{
> +	if ((addr >= mmu->mem_start) && (addr < (mmu->mem_start + mmu->mem_size)))
> +		return 1;
> +
> +	return 0;
> +}
> +
> +static inline struct mmu_info *mmu_lookup(u32 addr)
> +{
> +	u32 i;
> +
> +	for (i = 0; i < ARRAY_SIZE(mmu_list); i++) {
> +		if (mmu_check_mem_range(mmu_list[i], addr))
> +			return mmu_list[i];
> +	}
> +
> +	return NULL;
> +}
> +
> +static inline u32 mmu_virt_to_phys(u32 reg, u32 va, u32 mask)
> +{
> +	return (reg & mask) | (va & (~mask));
> +}
> +
> +static inline u32 mmu_phys_to_virt(u32 reg, u32 pa, u32 mask)
> +{
> +	return (reg & ~mask) | pa;
> +}
> +
> +static int mmu_mmio_check(struct vcpu *v, paddr_t addr)
> +{
> +	return mmu_for_each(mmu_check_mem_range, addr);
> +}
> +
> +static int mmu_copy_pagetable(struct mmu_info *mmu)
> +{
> +	void __iomem *pagetable = NULL;
> +	u32 pgaddr;
> +
> +	ASSERT(mmu);
> +
> +	/* read address where kernel MMU pagetable is stored */
> +	pgaddr = readl(mmu->mem_map + MMU_TTB);
> +	pagetable = ioremap(pgaddr, IOPGD_TABLE_SIZE);
> +	if (!pagetable) {

Xen uses a different coding style from Linux, see CODING_STYLE.


> +		printk("%s: %s failed to map pagetable\n",
> +			   __func__, mmu->name);
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * pagetable can be changed since last time
> +	 * we accessed it therefore we need to copy it each time
> +	 */
> +	memcpy(mmu->pagetable, pagetable, IOPGD_TABLE_SIZE);
> +
> +	iounmap(pagetable);

Do you need to flush the dcache here?


> +	return 0;
> +}
> +
> +#define mmu_dump_pdentry(da, iopgd, paddr, maddr, vaddr, mask)									\
> +{																								\
> +	const char *sect_type = (iopgd_is_table(iopgd) || (mask == IOPTE_SMALL_MASK) ||				\
> +							(mask == IOPTE_LARGE_MASK)) ? "table"								\
> +							: iopgd_is_super(iopgd) ? "supersection"							\
> +							: iopgd_is_section(iopgd) ? "section"								\
> +							: "Unknown section";												\
> +	printk("[iopgd] %s da 0x%08x iopgd 0x%08x paddr 0x%08x maddr 0x%pS vaddr 0x%08x mask 0x%08x\n",\
> +		   sect_type, da, iopgd, paddr, _p(maddr), vaddr, mask);								\
> +}
> +
> +static u32 mmu_translate_pgentry(struct domain *dom, u32 iopgd, u32 da, u32 mask)
> +{
> +	u32 vaddr, paddr;
> +	paddr_t maddr;
> +
> +	paddr = mmu_virt_to_phys(iopgd, da, mask);
> +	maddr = p2m_lookup(dom, paddr);
> +	vaddr = mmu_phys_to_virt(iopgd, maddr, mask);
> +
> +	return vaddr;
> +}
> +
> +/*
> + * on boot table is empty
> + */
> +static int mmu_translate_pagetable(struct domain *dom, struct mmu_info *mmu)
> +{
> +	u32 i;
> +	int res;
> +	bool table_updated = false;
> +
> +	ASSERT(dom);
> +	ASSERT(mmu);
> +
> +	/* copy pagetable from  domain to xen */
> +	res = mmu_copy_pagetable(mmu);
> +	if (res) {
> +		printk("%s: %s failed to map pagetable memory\n",
> +			   __func__, mmu->name);
> +		return res;
> +	}
> +
> +	/* 1-st level translation */
> +	for (i = 0; i < PTRS_PER_IOPGD; i++) {
> +		u32 da;
> +		u32 iopgd = mmu->pagetable[i];
> +
> +		if (!iopgd)
> +			continue;
> +
> +		table_updated = true;
> +
> +		/* "supersection" 16 Mb */
> +		if (iopgd_is_super(iopgd)) {
> +			da = i << IOSECTION_SHIFT;
> +			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSUPER_MASK);
> +
> +		/* "section" 1Mb */
> +		} else if (iopgd_is_section(iopgd)) {
> +			da = i << IOSECTION_SHIFT;
> +			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
> +
> +		/* "table" */
> +		} else if (iopgd_is_table(iopgd)) {
> +			u32 j, mask;
> +			u32 iopte = iopte_offset(iopgd);
> +
> +			/* 2-nd level translation */
> +			for (j = 0; j < PTRS_PER_IOPTE; j++, iopte += IOPTE_SMALL_SIZE) {
> +
> +				/* "small table" 4Kb */
> +				if (iopte_is_small(iopgd)) {
> +					da = (i << IOSECTION_SHIFT) + (j << IOPTE_SMALL_SHIFT);
> +					mask = IOPTE_SMALL_MASK;
> +
> +				/* "large table" 64Kb */
> +				} else if (iopte_is_large(iopgd)) {
> +					da = (i << IOSECTION_SHIFT) + (j << IOPTE_LARGE_SHIFT);
> +					mask = IOPTE_LARGE_MASK;
> +
> +				/* error */
> +				} else {
> +					printk("%s Unknown table type 0x%08x\n", mmu->name, iopte);
> +					return -EINVAL;
> +				}
> +
> +				/* translate 2-nd level entry */
> +				mmu->pagetable[i] = mmu_translate_pgentry(dom, iopte, da, mask);
> +			}
> +
> +			continue;
> +
> +		/* error */
> +		} else {
> +			printk("%s Unknown entry 0x%08x\n", mmu->name, iopgd);
> +			return -EINVAL;
> +		}
> +	}
> +
> +	/* force omap IOMMU to use new pagetable */
> +	if (table_updated) {
> +		paddr_t maddr;
> +		flush_xen_dcache_va_range(mmu->pagetable, IOPGD_TABLE_SIZE);

So you are flushing the dcache all at once at the end, probably better
this way.

> +		maddr = __pa(mmu->pagetable);
> +		writel(maddr, mmu->mem_map + MMU_TTB);
> +		printk("%s update pagetable, maddr 0x%pS\n", mmu->name, _p(maddr));
> +	}
> +
> +	return 0;
> +}
> +
> +static int mmu_trap_write_access(struct domain *dom,
> +								 struct mmu_info *mmu, mmio_info_t *info)
> +{
> +	struct cpu_user_regs *regs = guest_cpu_user_regs();
> +	register_t *r = select_user_reg(regs, info->dabt.reg);
> +	int res = 0;
> +
> +	switch (info->gpa - mmu->mem_start) {
> +		case MMU_TTB:
> +			printk("%s MMU_TTB write access 0x%pS <= 0x%08x\n",
> +				   mmu->name, _p(info->gpa), *r);
> +			res = mmu_translate_pagetable(dom, mmu);
> +			break;
> +		default:
> +			break;
> +	}
> +
> +	return res;
> +}
> +
> +static int mmu_mmio_read(struct vcpu *v, mmio_info_t *info)
> +{
> +	struct mmu_info *mmu = NULL;
> +    struct cpu_user_regs *regs = guest_cpu_user_regs();
> +    register_t *r = select_user_reg(regs, info->dabt.reg);
> +
> +	mmu = mmu_lookup(info->gpa);
> +	if (!mmu) {
> +		printk("%s: can't get mmu for addr 0x%08x\n", __func__, (u32)info->gpa);
> +		return -EINVAL;
> +	}
> +
> +    *r = readl(mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
> +
> +    return 1;
> +}
> +
> +static int mmu_mmio_write(struct vcpu *v, mmio_info_t *info)
> +{
> +	struct domain *dom = v->domain;
> +	struct mmu_info *mmu = NULL;
> +    struct cpu_user_regs *regs = guest_cpu_user_regs();
> +    register_t *r = select_user_reg(regs, info->dabt.reg);
> +	int res;
> +
> +	mmu = mmu_lookup(info->gpa);
> +	if (!mmu) {
> +		printk("%s: can't get mmu for addr 0x%08x\n", __func__, (u32)info->gpa);
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * make sure that user register is written first in this function
> +	 * following calls may expect valid data in it
> +	 */
> +    writel(*r, mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
> +
> +	res = mmu_trap_write_access(dom, mmu, info);
> +	if (res)
> +		return res;
> +
> +    return 1;
> +}

I wonder if we actually need to trap guest accesses in all cases: if we
leave the GPU/IPU in Dom0, mapped 1:1, then we don't need any traps.
Maybe we can find a way to detect that so that we can avoid trapping and
translating in that case.


> +static int mmu_init(struct mmu_info *mmu, u32 data)
> +{
> +	ASSERT(mmu);
> +	ASSERT(!mmu->mem_map);
> +	ASSERT(!mmu->pagetable);
> +
> +    mmu->mem_map = ioremap(mmu->mem_start, mmu->mem_size);
> +	if (!mmu->mem_map) {
> +		printk("%s: %s failed to map memory\n",  __func__, mmu->name);
> +		return -EINVAL;
> +	}
> +
> +	printk("%s: %s ipu_map = 0x%pS\n", __func__, mmu->name, _p(mmu->mem_map));
> +
> +	mmu->pagetable = xzalloc_bytes(IOPGD_TABLE_SIZE);
> +	if (!mmu->pagetable) {
> +		printk("%s: %s failed to alloc private pagetable\n",
> +			   __func__, mmu->name);
> +		return -ENOMEM;
> +	}
> +
> +	printk("%s: %s private pagetable %lu bytes\n",
> +		   __func__, mmu->name, IOPGD_TABLE_SIZE);
> +
> +	return 0;
> +}
> +
> +static int mmu_init_all(void)
> +{
> +	int res;
> +
> +	res = mmu_for_each(mmu_init, 0);
> +	if (res) {
> +		printk("%s error during init %d\n", __func__, res);
> +		return res;
> +	}
> +
> +	return 0;
> +}
> +
> +const struct mmio_handler mmu_mmio_handler = {
> +	.check_handler = mmu_mmio_check,
> +	.read_handler  = mmu_mmio_read,
> +	.write_handler = mmu_mmio_write,
> +};
> +
> +__initcall(mmu_init_all);
> -- 
> 1.7.9.5
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel
> 

Re: [RFC v01 1/3] arm: omap: introduce iommu module

[Andrii Tseglytskyi]

Hi Stefano,

On Thu, Jan 23, 2014 at 4:39 PM, Stefano Stabellini
<stefano.stabellini@eu.citrix.com> wrote:
> On Wed, 22 Jan 2014, Andrii Tseglytskyi wrote:
>> omap IOMMU module is designed to handle access to external
>> omap MMUs, connected to the L3 bus.
>>
...
>> @@ -26,6 +26,7 @@ static const struct mmio_handler *const mmio_handlers[] =
>>  {
>>      &vgic_distr_mmio_handler,
>>      &vuart_mmio_handler,
>> +     &mmu_mmio_handler,
>>  };
>>  #define MMIO_HANDLER_NR ARRAY_SIZE(mmio_handlers)
>
> I think that omap_iommu should be a platform specific driver, and it
> should hook into a set of platform specific mmio_handlers instead of
> using the generic mmio_handler structure.
>

Agree.

...

>> diff --git a/xen/arch/arm/io.h b/xen/arch/arm/io.h
>> index 8d252c0..acb5dff 100644
>> --- a/xen/arch/arm/io.h
>> +++ b/xen/arch/arm/io.h
>> @@ -42,6 +42,7 @@ struct mmio_handler {
>>
>>  extern const struct mmio_handler vgic_distr_mmio_handler;
>>  extern const struct mmio_handler vuart_mmio_handler;
>> +extern const struct mmio_handler mmu_mmio_handler;
>>
>>  extern int handle_mmio(mmio_info_t *info);
>>
>> diff --git a/xen/arch/arm/omap_iommu.c b/xen/arch/arm/omap_iommu.c
>> new file mode 100644
>> index 0000000..4dab30f
>> --- /dev/null
>> +++ b/xen/arch/arm/omap_iommu.c
>
> It should probably live under xen/arch/arm/platforms.
>

Agree.

...

>> +static int mmu_copy_pagetable(struct mmu_info *mmu)
>> +{
>> +     void __iomem *pagetable = NULL;
>> +     u32 pgaddr;
>> +
>> +     ASSERT(mmu);
>> +
>> +     /* read address where kernel MMU pagetable is stored */
>> +     pgaddr = readl(mmu->mem_map + MMU_TTB);
>> +     pagetable = ioremap(pgaddr, IOPGD_TABLE_SIZE);
>> +     if (!pagetable) {
>
> Xen uses a different coding style from Linux, see CODING_STYLE.
>

Sure. Thank you. Will update my editor settings accordingly to fit Xen
coding style.

...
>
>> +             printk("%s: %s failed to map pagetable\n",
>> +                        __func__, mmu->name);
>> +             return -EINVAL;
>> +     }
>> +
>> +     /*
>> +      * pagetable can be changed since last time
>> +      * we accessed it therefore we need to copy it each time
>> +      */
>> +     memcpy(mmu->pagetable, pagetable, IOPGD_TABLE_SIZE);
>> +
>> +     iounmap(pagetable);
>
> Do you need to flush the dcache here?
>

No, it is copying from kernel to Xen. Kernel already has valid
pagetable in it physical memory, when this call happens. Kernel makes
sure of this before accessing MMU configuration register. The only
reason to flush cache here if kernel mmu driver will be modified
somehow. This may be the point.

...

>> +
>> +             /* error */
>> +             } else {
>> +                     printk("%s Unknown entry 0x%08x\n", mmu->name, iopgd);
>> +                     return -EINVAL;
>> +             }
>> +     }
>> +
>> +     /* force omap IOMMU to use new pagetable */
>> +     if (table_updated) {
>> +             paddr_t maddr;
>> +             flush_xen_dcache_va_range(mmu->pagetable, IOPGD_TABLE_SIZE);
>
> So you are flushing the dcache all at once at the end, probably better
> this way.

Right. After all translations complete I flush caches. This guarantees
that phys memory will contain valid data for MMU.

 ...

>> +    writel(*r, mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
>> +
>> +     res = mmu_trap_write_access(dom, mmu, info);
>> +     if (res)
>> +             return res;
>> +
>> +    return 1;
>> +}
>
> I wonder if we actually need to trap guest accesses in all cases: if we
> leave the GPU/IPU in Dom0, mapped 1:1, then we don't need any traps.
> Maybe we can find a way to detect that so that we can avoid trapping and
> translating in that case.
>

If we leave the GPU/IPU in Dom0 with 1:1 mapping we won't need any
translation. But for now we need this in DomU, where 1:1 mapping will
not be available.

>> +
>> +const struct mmio_handler mmu_mmio_handler = {
>> +     .check_handler = mmu_mmio_check,
>> +     .read_handler  = mmu_mmio_read,
>> +     .write_handler = mmu_mmio_write,
>> +};
>> +
>> +__initcall(mmu_init_all);
>> --
>> 1.7.9.5
>>
>>
>> _______________________________________________
>> Xen-devel mailing list
>> Xen-devel@lists.xen.org
>> http://lists.xen.org/xen-devel
>>

Thank you for review,

Regards,
Andrii

-- 

Andrii Tseglytskyi | Embedded Dev
GlobalLogic
www.globallogic.com

Re: [RFC v01 1/3] arm: omap: introduce iommu module

[Stefano Stabellini]

On Fri, 24 Jan 2014, Andrii Tseglytskyi wrote:
> >> +    writel(*r, mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
> >> +
> >> +     res = mmu_trap_write_access(dom, mmu, info);
> >> +     if (res)
> >> +             return res;
> >> +
> >> +    return 1;
> >> +}
> >
> > I wonder if we actually need to trap guest accesses in all cases: if we
> > leave the GPU/IPU in Dom0, mapped 1:1, then we don't need any traps.
> > Maybe we can find a way to detect that so that we can avoid trapping and
> > translating in that case.
> >
> 
> If we leave the GPU/IPU in Dom0 with 1:1 mapping we won't need any
> translation. But for now we need this in DomU, where 1:1 mapping will
> not be available.

Sure, that is fine.

What I meant to say is that in order to make this code more generic, it
would be a good idea to check whether the guest is dom0 or domU and only
trap MMU accesses if the guest is domU (because for dom0 is not
necessary).

Re: [RFC v01 1/3] arm: omap: introduce iommu module

[Andrii Tseglytskyi]

Hi,


On Fri, Jan 24, 2014 at 1:49 PM, Stefano Stabellini
<stefano.stabellini@eu.citrix.com> wrote:
> On Fri, 24 Jan 2014, Andrii Tseglytskyi wrote:
>>
>> If we leave the GPU/IPU in Dom0 with 1:1 mapping we won't need any
>> translation. But for now we need this in DomU, where 1:1 mapping will
>> not be available.
>
> Sure, that is fine.
>
> What I meant to say is that in order to make this code more generic, it
> would be a good idea to check whether the guest is dom0 or domU and only
> trap MMU accesses if the guest is domU (because for dom0 is not
> necessary).

Oh, got it. Sure will make this check.


regards,
Andrii

-- 

Andrii Tseglytskyi | Embedded Dev
GlobalLogic
www.globallogic.com

Re: [RFC v01 1/3] arm: omap: introduce iommu module

[Julien Grall]

On 01/22/2014 03:52 PM, Andrii Tseglytskyi wrote:
> omap IOMMU module is designed to handle access to external
> omap MMUs, connected to the L3 bus.

Hello Andrii,

Thanks for the patch. See my comment inline (I won't add the same
comment as Stefano).

> Change-Id: I96bbf2738e9dd2e21662e0986ca15c60183e669e
> Signed-off-by: Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
> ---

[..]

> +struct mmu_info {
> +	const char			*name;
> +	paddr_t				mem_start;
> +	u32					mem_size;
> +	u32					*pagetable;
> +	void __iomem		*mem_map;
> +};
> +
> +static struct mmu_info omap_ipu_mmu = {

static const?

> +	.name		= "IPU_L2_MMU",
> +	.mem_start	= 0x55082000,
> +	.mem_size	= 0x1000,
> +	.pagetable	= NULL,
> +};
> +
> +static struct mmu_info omap_dsp_mmu = {

static const?

> +	.name		= "DSP_L2_MMU",
> +	.mem_start	= 0x4a066000,
> +	.mem_size	= 0x1000,
> +	.pagetable	= NULL,
> +};
> +
> +static struct mmu_info *mmu_list[] = {

static const?

> +	&omap_ipu_mmu,
> +	&omap_dsp_mmu,
> +};
> +
> +#define mmu_for_each(pfunc, data)						\
> +({														\
> +	u32 __i;											\
> +	int __res = 0;										\
> +														\
> +	for (__i = 0; __i < ARRAY_SIZE(mmu_list); __i++) {	\
> +		__res |= pfunc(mmu_list[__i], data);			\

You res |= will result to a "wrong" errno if you have multiple failure.
Would it be better to have:

__res = pfunc(...)
if ( __res )
  break;

> +	}													\
> +	__res;												\
> +})
> +
> +static int mmu_check_mem_range(struct mmu_info *mmu, paddr_t addr)
> +{
> +	if ((addr >= mmu->mem_start) && (addr < (mmu->mem_start + mmu->mem_size)))
> +		return 1;
> +
> +	return 0;
> +}
> +
> +static inline struct mmu_info *mmu_lookup(u32 addr)
> +{
> +	u32 i;
> +
> +	for (i = 0; i < ARRAY_SIZE(mmu_list); i++) {
> +		if (mmu_check_mem_range(mmu_list[i], addr))
> +			return mmu_list[i];
> +	}
> +
> +	return NULL;
> +}
> +
> +static inline u32 mmu_virt_to_phys(u32 reg, u32 va, u32 mask)
> +{
> +	return (reg & mask) | (va & (~mask));
> +}
> +
> +static inline u32 mmu_phys_to_virt(u32 reg, u32 pa, u32 mask)
> +{
> +	return (reg & ~mask) | pa;
> +}
> +
> +static int mmu_mmio_check(struct vcpu *v, paddr_t addr)
> +{
> +	return mmu_for_each(mmu_check_mem_range, addr);
> +}

As I understand your cover letter, the device (and therefore the MMU) is
only passthrough to a single guest, right?

If so, your mmu_mmio_check should check if the domain is handling the
device.
With your current code any guest can write to this range and rewriting
the MMU page table.

> +
> +static int mmu_copy_pagetable(struct mmu_info *mmu)
> +{
> +	void __iomem *pagetable = NULL;
> +	u32 pgaddr;
> +
> +	ASSERT(mmu);
> +
> +	/* read address where kernel MMU pagetable is stored */
> +	pgaddr = readl(mmu->mem_map + MMU_TTB);
> +	pagetable = ioremap(pgaddr, IOPGD_TABLE_SIZE);
> +	if (!pagetable) {
> +		printk("%s: %s failed to map pagetable\n",
> +			   __func__, mmu->name);
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * pagetable can be changed since last time
> +	 * we accessed it therefore we need to copy it each time
> +	 */
> +	memcpy(mmu->pagetable, pagetable, IOPGD_TABLE_SIZE);
> +
> +	iounmap(pagetable);
> +
> +	return 0;
> +}

I'm confused, it should copy from the guest MMU pagetable, right? In
this case you should use map_domain_page.
ioremap *MUST* only be used with device memory, otherwise memory
coherency is not guaranteed.

[..]

> +static int mmu_mmio_write(struct vcpu *v, mmio_info_t *info)
> +{
> +	struct domain *dom = v->domain;
> +	struct mmu_info *mmu = NULL;
> +    struct cpu_user_regs *regs = guest_cpu_user_regs();
> +    register_t *r = select_user_reg(regs, info->dabt.reg);
> +	int res;
> +
> +	mmu = mmu_lookup(info->gpa);
> +	if (!mmu) {
> +		printk("%s: can't get mmu for addr 0x%08x\n", __func__, (u32)info->gpa);
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * make sure that user register is written first in this function
> +	 * following calls may expect valid data in it
> +	 */
> +    writel(*r, mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));

Hmmm ... I think this is very confusing, you should only write to the
memory if mmu_trap_write_access has not failed. And use "*r" where it's
needed.

Writing to the device memory could have side effect (for instance
updating the page table with the wrong translation...).

> +
> +	res = mmu_trap_write_access(dom, mmu, info);
> +	if (res)
> +		return res;
> +
> +    return 1;
> +}
> +
> +static int mmu_init(struct mmu_info *mmu, u32 data)
> +{
> +	ASSERT(mmu);
> +	ASSERT(!mmu->mem_map);
> +	ASSERT(!mmu->pagetable);
> +
> +    mmu->mem_map = ioremap(mmu->mem_start, mmu->mem_size);

Can you use ioremap_nocache instead of ioremap? The behavior is the same
but the former name is less confusing.

-- 
Julien Grall

Re: [RFC v01 1/3] arm: omap: introduce iommu module

[Andrii Tseglytskyi]

Hi Julien,

On Thu, Jan 23, 2014 at 5:31 PM, Julien Grall <julien.grall@linaro.org> wrote:
> On 01/22/2014 03:52 PM, Andrii Tseglytskyi wrote:
>> omap IOMMU module is designed to handle access to external
>> omap MMUs, connected to the L3 bus.

[...]

>
>> +struct mmu_info {
>> +     const char                      *name;
>> +     paddr_t                         mem_start;
>> +     u32                                     mem_size;
>> +     u32                                     *pagetable;
>> +     void __iomem            *mem_map;
>> +};
>> +
>> +static struct mmu_info omap_ipu_mmu = {
>
> static const?
>

Unfortunately, no. I like const modifiers very much and try to put
them everywhere I can, but in these structs I need to modify several
fields during MMU configuratiion.

[...]

>> +     .name           = "IPU_L2_MMU",
>> +     .mem_start      = 0x55082000,
>> +     .mem_size       = 0x1000,
>> +     .pagetable      = NULL,
>> +};
>> +
>> +static struct mmu_info omap_dsp_mmu = {
>
> static const?
>

The same as previous.

>> +     .name           = "DSP_L2_MMU",
>> +     .mem_start      = 0x4a066000,
>> +     .mem_size       = 0x1000,
>> +     .pagetable      = NULL,
>> +};
>> +
>> +static struct mmu_info *mmu_list[] = {
>
> static const?
>

The same as previous.

>> +     &omap_ipu_mmu,
>> +     &omap_dsp_mmu,
>> +};
>> +
>> +#define mmu_for_each(pfunc, data)                                            \
>> +({                                                                                                           \
>> +     u32 __i;                                                                                        \
>> +     int __res = 0;                                                                          \
>> +                                                                                                             \
>> +     for (__i = 0; __i < ARRAY_SIZE(mmu_list); __i++) {      \
>> +             __res |= pfunc(mmu_list[__i], data);                    \
>
> You res |= will result to a "wrong" errno if you have multiple failure.
> Would it be better to have:
>
> __res = pfunc(...)
> if ( __res )
>   break;
>

I know. I tried both solutions - mine and what you proposed. Agree in
general, will update this.


>> +     }                                                                                                       \
>> +     __res;                                                                                          \
>> +})
>> +
>> +static int mmu_check_mem_range(struct mmu_info *mmu, paddr_t addr)
>> +{
>> +     if ((addr >= mmu->mem_start) && (addr < (mmu->mem_start + mmu->mem_size)))
>> +             return 1;
>> +
>> +     return 0;
>> +}
>> +
>> +static inline struct mmu_info *mmu_lookup(u32 addr)
>> +{
>> +     u32 i;
>> +
>> +     for (i = 0; i < ARRAY_SIZE(mmu_list); i++) {
>> +             if (mmu_check_mem_range(mmu_list[i], addr))
>> +                     return mmu_list[i];
>> +     }
>> +
>> +     return NULL;
>> +}
>> +
>> +static inline u32 mmu_virt_to_phys(u32 reg, u32 va, u32 mask)
>> +{
>> +     return (reg & mask) | (va & (~mask));
>> +}
>> +
>> +static inline u32 mmu_phys_to_virt(u32 reg, u32 pa, u32 mask)
>> +{
>> +     return (reg & ~mask) | pa;
>> +}
>> +
>> +static int mmu_mmio_check(struct vcpu *v, paddr_t addr)
>> +{
>> +     return mmu_for_each(mmu_check_mem_range, addr);
>> +}
>
> As I understand your cover letter, the device (and therefore the MMU) is
> only passthrough to a single guest, right?
>
> If so, your mmu_mmio_check should check if the domain is handling the
> device.
> With your current code any guest can write to this range and rewriting
> the MMU page table.
>

Oh, I knew that someone will catch this :)
This is a next step for this patch series - to make sure that only one
guest can configure / access MMU.

>> +
>> +static int mmu_copy_pagetable(struct mmu_info *mmu)
>> +{
>> +     void __iomem *pagetable = NULL;
>> +     u32 pgaddr;
>> +
>> +     ASSERT(mmu);
>> +
>> +     /* read address where kernel MMU pagetable is stored */
>> +     pgaddr = readl(mmu->mem_map + MMU_TTB);
>> +     pagetable = ioremap(pgaddr, IOPGD_TABLE_SIZE);
>> +     if (!pagetable) {
>> +             printk("%s: %s failed to map pagetable\n",
>> +                        __func__, mmu->name);
>> +             return -EINVAL;
>> +     }
>> +
>> +     /*
>> +      * pagetable can be changed since last time
>> +      * we accessed it therefore we need to copy it each time
>> +      */
>> +     memcpy(mmu->pagetable, pagetable, IOPGD_TABLE_SIZE);
>> +
>> +     iounmap(pagetable);
>> +
>> +     return 0;
>> +}
>
> I'm confused, it should copy from the guest MMU pagetable, right? In
> this case you should use map_domain_page.
> ioremap *MUST* only be used with device memory, otherwise memory
> coherency is not guaranteed.
>

OK. Will try this.

> [..]
>
>> +static int mmu_mmio_write(struct vcpu *v, mmio_info_t *info)
>> +{
>> +     struct domain *dom = v->domain;
>> +     struct mmu_info *mmu = NULL;
>> +    struct cpu_user_regs *regs = guest_cpu_user_regs();
>> +    register_t *r = select_user_reg(regs, info->dabt.reg);
>> +     int res;
>> +
>> +     mmu = mmu_lookup(info->gpa);
>> +     if (!mmu) {
>> +             printk("%s: can't get mmu for addr 0x%08x\n", __func__, (u32)info->gpa);
>> +             return -EINVAL;
>> +     }
>> +
>> +     /*
>> +      * make sure that user register is written first in this function
>> +      * following calls may expect valid data in it
>> +      */
>> +    writel(*r, mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
>
> Hmmm ... I think this is very confusing, you should only write to the
> memory if mmu_trap_write_access has not failed. And use "*r" where it's
> needed.
>
> Writing to the device memory could have side effect (for instance
> updating the page table with the wrong translation...).
>

Agree - it is a bit confusing here. But I need a valid data in the
user register.
Following calls use it ->
mmu_trap_write_access()->mmu_translate_pagetable()->mmu_copy_pagetable()->pgaddr
= readl(mmu->mem_map + MMU_TTB);
Last read will be from register written in this function. Taking in
account your comment - I will think about changing this logic.

>> +
>> +     res = mmu_trap_write_access(dom, mmu, info);
>> +     if (res)
>> +             return res;
>> +
>> +    return 1;
>> +}
>> +
>> +static int mmu_init(struct mmu_info *mmu, u32 data)
>> +{
>> +     ASSERT(mmu);
>> +     ASSERT(!mmu->mem_map);
>> +     ASSERT(!mmu->pagetable);
>> +
>> +    mmu->mem_map = ioremap(mmu->mem_start, mmu->mem_size);
>
> Can you use ioremap_nocache instead of ioremap? The behavior is the same
> but the former name is less confusing.
>

Sure.

> --
> Julien Grall

Thank you for review.

Regards,
Andrii


-- 

Andrii Tseglytskyi | Embedded Dev
GlobalLogic
www.globallogic.com

Re: [RFC v01 1/3] arm: omap: introduce iommu module

[Julien Grall]

On 01/24/2014 11:49 AM, Andrii Tseglytskyi wrote:
> Hi Julien,
> 
> On Thu, Jan 23, 2014 at 5:31 PM, Julien Grall <julien.grall@linaro.org> wrote:
>> On 01/22/2014 03:52 PM, Andrii Tseglytskyi wrote:
>>> omap IOMMU module is designed to handle access to external
>>> omap MMUs, connected to the L3 bus.
> 
> [...]
> 
>>
>>> +struct mmu_info {
>>> +     const char                      *name;
>>> +     paddr_t                         mem_start;
>>> +     u32                                     mem_size;
>>> +     u32                                     *pagetable;
>>> +     void __iomem            *mem_map;
>>> +};
>>> +
>>> +static struct mmu_info omap_ipu_mmu = {
>>
>> static const?
>>
> 
> Unfortunately, no. I like const modifiers very much and try to put
> them everywhere I can, but in these structs I need to modify several
> fields during MMU configuratiion.
> 
> [...]
> 
>>> +     .name           = "IPU_L2_MMU",
>>> +     .mem_start      = 0x55082000,
>>> +     .mem_size       = 0x1000,
>>> +     .pagetable      = NULL,
>>> +};
>>> +
>>> +static struct mmu_info omap_dsp_mmu = {
>>
>> static const?
>>
> 
> The same as previous.
> 
>>> +     .name           = "DSP_L2_MMU",
>>> +     .mem_start      = 0x4a066000,
>>> +     .mem_size       = 0x1000,
>>> +     .pagetable      = NULL,
>>> +};
>>> +
>>> +static struct mmu_info *mmu_list[] = {
>>
>> static const?
>>
> 
> The same as previous.
> 
>>> +     &omap_ipu_mmu,
>>> +     &omap_dsp_mmu,
>>> +};
>>> +
>>> +#define mmu_for_each(pfunc, data)                                            \
>>> +({                                                                                                           \
>>> +     u32 __i;                                                                                        \
>>> +     int __res = 0;                                                                          \
>>> +                                                                                                             \
>>> +     for (__i = 0; __i < ARRAY_SIZE(mmu_list); __i++) {      \
>>> +             __res |= pfunc(mmu_list[__i], data);                    \
>>
>> You res |= will result to a "wrong" errno if you have multiple failure.
>> Would it be better to have:
>>
>> __res = pfunc(...)
>> if ( __res )
>>   break;
>>
> 
> I know. I tried both solutions - mine and what you proposed. Agree in
> general, will update this.
> 
> 
>>> +     }                                                                                                       \
>>> +     __res;                                                                                          \
>>> +})
>>> +
>>> +static int mmu_check_mem_range(struct mmu_info *mmu, paddr_t addr)
>>> +{
>>> +     if ((addr >= mmu->mem_start) && (addr < (mmu->mem_start + mmu->mem_size)))
>>> +             return 1;
>>> +
>>> +     return 0;
>>> +}
>>> +
>>> +static inline struct mmu_info *mmu_lookup(u32 addr)
>>> +{
>>> +     u32 i;
>>> +
>>> +     for (i = 0; i < ARRAY_SIZE(mmu_list); i++) {
>>> +             if (mmu_check_mem_range(mmu_list[i], addr))
>>> +                     return mmu_list[i];
>>> +     }
>>> +
>>> +     return NULL;
>>> +}
>>> +
>>> +static inline u32 mmu_virt_to_phys(u32 reg, u32 va, u32 mask)
>>> +{
>>> +     return (reg & mask) | (va & (~mask));
>>> +}
>>> +
>>> +static inline u32 mmu_phys_to_virt(u32 reg, u32 pa, u32 mask)
>>> +{
>>> +     return (reg & ~mask) | pa;
>>> +}
>>> +
>>> +static int mmu_mmio_check(struct vcpu *v, paddr_t addr)
>>> +{
>>> +     return mmu_for_each(mmu_check_mem_range, addr);
>>> +}
>>
>> As I understand your cover letter, the device (and therefore the MMU) is
>> only passthrough to a single guest, right?
>>
>> If so, your mmu_mmio_check should check if the domain is handling the
>> device.
>> With your current code any guest can write to this range and rewriting
>> the MMU page table.
>>
> 
> Oh, I knew that someone will catch this :)
> This is a next step for this patch series - to make sure that only one
> guest can configure / access MMU.
> 
>>> +
>>> +static int mmu_copy_pagetable(struct mmu_info *mmu)
>>> +{
>>> +     void __iomem *pagetable = NULL;
>>> +     u32 pgaddr;
>>> +
>>> +     ASSERT(mmu);
>>> +
>>> +     /* read address where kernel MMU pagetable is stored */
>>> +     pgaddr = readl(mmu->mem_map + MMU_TTB);
>>> +     pagetable = ioremap(pgaddr, IOPGD_TABLE_SIZE);
>>> +     if (!pagetable) {
>>> +             printk("%s: %s failed to map pagetable\n",
>>> +                        __func__, mmu->name);
>>> +             return -EINVAL;
>>> +     }
>>> +
>>> +     /*
>>> +      * pagetable can be changed since last time
>>> +      * we accessed it therefore we need to copy it each time
>>> +      */
>>> +     memcpy(mmu->pagetable, pagetable, IOPGD_TABLE_SIZE);
>>> +
>>> +     iounmap(pagetable);
>>> +
>>> +     return 0;
>>> +}
>>
>> I'm confused, it should copy from the guest MMU pagetable, right? In
>> this case you should use map_domain_page.
>> ioremap *MUST* only be used with device memory, otherwise memory
>> coherency is not guaranteed.
>>
> 
> OK. Will try this.

You can look at to __copy_{to,from}* macro. They will do the right job.

> 
>> [..]
>>
>>> +static int mmu_mmio_write(struct vcpu *v, mmio_info_t *info)
>>> +{
>>> +     struct domain *dom = v->domain;
>>> +     struct mmu_info *mmu = NULL;
>>> +    struct cpu_user_regs *regs = guest_cpu_user_regs();
>>> +    register_t *r = select_user_reg(regs, info->dabt.reg);
>>> +     int res;
>>> +
>>> +     mmu = mmu_lookup(info->gpa);
>>> +     if (!mmu) {
>>> +             printk("%s: can't get mmu for addr 0x%08x\n", __func__, (u32)info->gpa);
>>> +             return -EINVAL;
>>> +     }
>>> +
>>> +     /*
>>> +      * make sure that user register is written first in this function
>>> +      * following calls may expect valid data in it
>>> +      */
>>> +    writel(*r, mmu->mem_map + ((u32)(info->gpa) - mmu->mem_start));
>>
>> Hmmm ... I think this is very confusing, you should only write to the
>> memory if mmu_trap_write_access has not failed. And use "*r" where it's
>> needed.
>>
>> Writing to the device memory could have side effect (for instance
>> updating the page table with the wrong translation...).
>>
> 
> Agree - it is a bit confusing here. But I need a valid data in the
> user register.
> Following calls use it ->
> mmu_trap_write_access()->mmu_translate_pagetable()->mmu_copy_pagetable()->pgaddr
> = readl(mmu->mem_map + MMU_TTB);
> Last read will be from register written in this function. Taking in
> account your comment - I will think about changing this logic.
> 
>>> +
>>> +     res = mmu_trap_write_access(dom, mmu, info);
>>> +     if (res)
>>> +             return res;
>>> +
>>> +    return 1;
>>> +}
>>> +
>>> +static int mmu_init(struct mmu_info *mmu, u32 data)
>>> +{
>>> +     ASSERT(mmu);
>>> +     ASSERT(!mmu->mem_map);
>>> +     ASSERT(!mmu->pagetable);
>>> +
>>> +    mmu->mem_map = ioremap(mmu->mem_start, mmu->mem_size);
>>
>> Can you use ioremap_nocache instead of ioremap? The behavior is the same
>> but the former name is less confusing.
>>
> 
> Sure.
> 
>> --
>> Julien Grall
> 
> Thank you for review.
> 
> Regards,
> Andrii
> 
> 


-- 
Julien Grall

[RFC v01 2/3] arm: omap: translate iommu mapping to 4K pages

[Andrii Tseglytskyi]

Patch introduces the following algorithm:
- enumerates all first level translation entries
- for each section creates 256 pages, each page is 4096 bytes
- for each supersection creates 4096 pages, each page is 4096 bytes
- flush cache to synchronize Cortex M15 and IOMMU

This algorithm make possible to use 4K mapping only.

Change-Id: Ie2cf45f23e0c170e9ba9d58f8dbb917348fdbd33
Signed-off-by: Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
---
 xen/arch/arm/omap_iommu.c |   50 +++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 46 insertions(+), 4 deletions(-)

diff --git a/xen/arch/arm/omap_iommu.c b/xen/arch/arm/omap_iommu.c
index 4dab30f..7ec03a2 100644
--- a/xen/arch/arm/omap_iommu.c
+++ b/xen/arch/arm/omap_iommu.c
@@ -72,6 +72,9 @@
 #define PTRS_PER_IOPTE		(1UL << (IOPGD_SHIFT - IOPTE_SMALL_SHIFT))
 #define IOPTE_TABLE_SIZE	(PTRS_PER_IOPTE * sizeof(u32))
 
+/* 16 sections in supersection */
+#define IOSECTION_PER_IOSUPER	(1UL << (IOSUPER_SHIFT - IOPGD_SHIFT))
+
 /*
  * some descriptor attributes.
  */
@@ -117,6 +120,9 @@ static struct mmu_info *mmu_list[] = {
 	&omap_dsp_mmu,
 };
 
+static bool translate_supersections_to_pages = true;
+static bool translate_sections_to_pages = true;
+
 #define mmu_for_each(pfunc, data)						\
 ({														\
 	u32 __i;											\
@@ -213,6 +219,29 @@ static u32 mmu_translate_pgentry(struct domain *dom, u32 iopgd, u32 da, u32 mask
 	return vaddr;
 }
 
+static u32 mmu_iopte_alloc(struct mmu_info *mmu, struct domain *dom, u32 iopgd, u32 sect_num)
+{
+	u32 *iopte = NULL;
+	u32 i;
+
+	iopte = xzalloc_bytes(PAGE_SIZE);
+	if (!iopte) {
+		printk("%s Fail to alloc 2nd level table\n", mmu->name);
+		return 0;
+	}
+
+	for (i = 0; i < PTRS_PER_IOPTE; i++) {
+		u32 da, vaddr, iopgd_tmp;
+		da = (sect_num << IOSECTION_SHIFT) + (i << IOPTE_SMALL_SHIFT);
+		iopgd_tmp = (iopgd & IOSECTION_MASK) + (i << IOPTE_SMALL_SHIFT);
+		vaddr = mmu_translate_pgentry(dom, iopgd_tmp, da, IOPTE_SMALL_MASK);
+		iopte[i] = vaddr | IOPTE_SMALL;
+	}
+
+	flush_xen_dcache_va_range(iopte, PAGE_SIZE);
+	return __pa(iopte) | IOPGD_TABLE;
+}
+
 /*
  * on boot table is empty
  */
@@ -245,13 +274,26 @@ static int mmu_translate_pagetable(struct domain *dom, struct mmu_info *mmu)
 
 		/* "supersection" 16 Mb */
 		if (iopgd_is_super(iopgd)) {
-			da = i << IOSECTION_SHIFT;
-			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSUPER_MASK);
+			if(likely(translate_supersections_to_pages)) {
+				u32 j, iopgd_tmp;
+				for (j = 0 ; j < IOSECTION_PER_IOSUPER; j++) {
+					iopgd_tmp = iopgd + (j * IOSECTION_SIZE);
+					mmu->pagetable[i + j] = mmu_iopte_alloc(mmu, dom, iopgd_tmp, i);
+				}
+				i += (j - 1);
+			} else {
+				da = i << IOSECTION_SHIFT;
+				mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSUPER_MASK);
+			}
 
 		/* "section" 1Mb */
 		} else if (iopgd_is_section(iopgd)) {
-			da = i << IOSECTION_SHIFT;
-			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
+			if (likely(translate_sections_to_pages)) {
+				mmu->pagetable[i] = mmu_iopte_alloc(mmu, dom, iopgd, i);
+			} else {
+				da = i << IOSECTION_SHIFT;
+				mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
+			}
 
 		/* "table" */
 		} else if (iopgd_is_table(iopgd)) {
-- 
1.7.9.5

Re: [RFC v01 2/3] arm: omap: translate iommu mapping to 4K pages

[Stefano Stabellini]

On Wed, 22 Jan 2014, Andrii Tseglytskyi wrote:
> Patch introduces the following algorithm:
> - enumerates all first level translation entries
> - for each section creates 256 pages, each page is 4096 bytes
> - for each supersection creates 4096 pages, each page is 4096 bytes
> - flush cache to synchronize Cortex M15 and IOMMU
> 
> This algorithm make possible to use 4K mapping only.
> 
> Change-Id: Ie2cf45f23e0c170e9ba9d58f8dbb917348fdbd33
> Signed-off-by: Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>

I take that the first patch doesn't actually work without this one?
In that case it might make sense to just merge them into one.


>  xen/arch/arm/omap_iommu.c |   50 +++++++++++++++++++++++++++++++++++++++++----
>  1 file changed, 46 insertions(+), 4 deletions(-)
> 
> diff --git a/xen/arch/arm/omap_iommu.c b/xen/arch/arm/omap_iommu.c
> index 4dab30f..7ec03a2 100644
> --- a/xen/arch/arm/omap_iommu.c
> +++ b/xen/arch/arm/omap_iommu.c
> @@ -72,6 +72,9 @@
>  #define PTRS_PER_IOPTE		(1UL << (IOPGD_SHIFT - IOPTE_SMALL_SHIFT))
>  #define IOPTE_TABLE_SIZE	(PTRS_PER_IOPTE * sizeof(u32))
>  
> +/* 16 sections in supersection */
> +#define IOSECTION_PER_IOSUPER	(1UL << (IOSUPER_SHIFT - IOPGD_SHIFT))
> +
>  /*
>   * some descriptor attributes.
>   */
> @@ -117,6 +120,9 @@ static struct mmu_info *mmu_list[] = {
>  	&omap_dsp_mmu,
>  };
>  
> +static bool translate_supersections_to_pages = true;
> +static bool translate_sections_to_pages = true;
> +
>  #define mmu_for_each(pfunc, data)						\
>  ({														\
>  	u32 __i;											\
> @@ -213,6 +219,29 @@ static u32 mmu_translate_pgentry(struct domain *dom, u32 iopgd, u32 da, u32 mask
>  	return vaddr;
>  }
>  
> +static u32 mmu_iopte_alloc(struct mmu_info *mmu, struct domain *dom, u32 iopgd, u32 sect_num)
> +{
> +	u32 *iopte = NULL;
> +	u32 i;
> +
> +	iopte = xzalloc_bytes(PAGE_SIZE);
> +	if (!iopte) {
> +		printk("%s Fail to alloc 2nd level table\n", mmu->name);
> +		return 0;
> +	}
> +
> +	for (i = 0; i < PTRS_PER_IOPTE; i++) {
> +		u32 da, vaddr, iopgd_tmp;
> +		da = (sect_num << IOSECTION_SHIFT) + (i << IOPTE_SMALL_SHIFT);
> +		iopgd_tmp = (iopgd & IOSECTION_MASK) + (i << IOPTE_SMALL_SHIFT);
> +		vaddr = mmu_translate_pgentry(dom, iopgd_tmp, da, IOPTE_SMALL_MASK);
> +		iopte[i] = vaddr | IOPTE_SMALL;
> +	}
> +
> +	flush_xen_dcache_va_range(iopte, PAGE_SIZE);
> +	return __pa(iopte) | IOPGD_TABLE;
> +}
> +
>  /*
>   * on boot table is empty
>   */
> @@ -245,13 +274,26 @@ static int mmu_translate_pagetable(struct domain *dom, struct mmu_info *mmu)
>  
>  		/* "supersection" 16 Mb */
>  		if (iopgd_is_super(iopgd)) {
> -			da = i << IOSECTION_SHIFT;
> -			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSUPER_MASK);
> +			if(likely(translate_supersections_to_pages)) {
> +				u32 j, iopgd_tmp;
> +				for (j = 0 ; j < IOSECTION_PER_IOSUPER; j++) {
> +					iopgd_tmp = iopgd + (j * IOSECTION_SIZE);
> +					mmu->pagetable[i + j] = mmu_iopte_alloc(mmu, dom, iopgd_tmp, i);
> +				}
> +				i += (j - 1);
> +			} else {
> +				da = i << IOSECTION_SHIFT;
> +				mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSUPER_MASK);
> +			}
>  
>  		/* "section" 1Mb */
>  		} else if (iopgd_is_section(iopgd)) {
> -			da = i << IOSECTION_SHIFT;
> -			mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
> +			if (likely(translate_sections_to_pages)) {
> +				mmu->pagetable[i] = mmu_iopte_alloc(mmu, dom, iopgd, i);
> +			} else {
> +				da = i << IOSECTION_SHIFT;
> +				mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
> +			}
>  
>  		/* "table" */
>  		} else if (iopgd_is_table(iopgd)) {

Since the 16MB and 1MB sections might not actually be contigous in
machine address space, this patch replaces the guest allocated sections
with pte tables pointing to the original IPAs. Is that right?

Re: [RFC v01 2/3] arm: omap: translate iommu mapping to 4K pages

[Andrii Tseglytskyi]

Hi Stefano,

On Thu, Jan 23, 2014 at 4:52 PM, Stefano Stabellini
<stefano.stabellini@eu.citrix.com> wrote:
> On Wed, 22 Jan 2014, Andrii Tseglytskyi wrote:
>> Patch introduces the following algorithm:
>> - enumerates all first level translation entries
>> - for each section creates 256 pages, each page is 4096 bytes
>> - for each supersection creates 4096 pages, each page is 4096 bytes
>> - flush cache to synchronize Cortex M15 and IOMMU
>>
>> This algorithm make possible to use 4K mapping only.
>>
>> Change-Id: Ie2cf45f23e0c170e9ba9d58f8dbb917348fdbd33
>> Signed-off-by: Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
>
> I take that the first patch doesn't actually work without this one?
> In that case it might make sense to just merge them into one.
>

No, each patch from the series works standalone. This patch rewrites
structure of the pagetable. I did this to increase granularity of
memory allocations. I'm pretty sure that we won't have continuous
blocks of 1Mb and 16 Mb
of memory, when several guest domains will be active. With 4K
granularity chances for successful translation are much more higher.

>>               /* "section" 1Mb */
>>               } else if (iopgd_is_section(iopgd)) {
>> -                     da = i << IOSECTION_SHIFT;
>> -                     mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
>> +                     if (likely(translate_sections_to_pages)) {
>> +                             mmu->pagetable[i] = mmu_iopte_alloc(mmu, dom, iopgd, i);
>> +                     } else {
>> +                             da = i << IOSECTION_SHIFT;
>> +                             mmu->pagetable[i] = mmu_translate_pgentry(dom, iopgd, da, IOSECTION_MASK);
>> +                     }
>>
>>               /* "table" */
>>               } else if (iopgd_is_table(iopgd)) {
>
> Since the 16MB and 1MB sections might not actually be contigous in
> machine address space, this patch replaces the guest allocated sections
> with pte tables pointing to the original IPAs. Is that right?

Yes, you are right.

Thank you for review,

regards,
Andrii

-- 

Andrii Tseglytskyi | Embedded Dev
GlobalLogic
www.globallogic.com

[RFC v01 3/3] arm: omap: cleanup iopte allocations

[Andrii Tseglytskyi]

Each allocation for iopte requires 4Kb memory.
All previous allocations from previous MMU reconfiguration
must be cleaned before new reconfigureation cycle.

Change-Id: I6db69a400cdba1170b43d9dc68d0817db77cbf9c
Signed-off-by: Andrii Tseglytskyi <andrii.tseglytskyi@globallogic.com>
---
 xen/arch/arm/omap_iommu.c |   35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

diff --git a/xen/arch/arm/omap_iommu.c b/xen/arch/arm/omap_iommu.c
index 7ec03a2..a5ad3ac 100644
--- a/xen/arch/arm/omap_iommu.c
+++ b/xen/arch/arm/omap_iommu.c
@@ -93,12 +93,18 @@
 #define iopte_is_large(x)	(((x) & 3) == IOPTE_LARGE)
 #define iopte_offset(x)		((x) & IOPTE_SMALL_MASK)
 
+struct mmu_alloc_node {
+	u32					*vptr;
+	struct list_head	node;
+};
+
 struct mmu_info {
 	const char			*name;
 	paddr_t				mem_start;
 	u32					mem_size;
 	u32					*pagetable;
 	void __iomem		*mem_map;
+	struct list_head	alloc_list;
 };
 
 static struct mmu_info omap_ipu_mmu = {
@@ -222,8 +228,15 @@ static u32 mmu_translate_pgentry(struct domain *dom, u32 iopgd, u32 da, u32 mask
 static u32 mmu_iopte_alloc(struct mmu_info *mmu, struct domain *dom, u32 iopgd, u32 sect_num)
 {
 	u32 *iopte = NULL;
+	struct mmu_alloc_node *alloc_node;
 	u32 i;
 
+	alloc_node = xzalloc_bytes(sizeof(struct mmu_alloc_node));
+	if (!alloc_node) {
+		printk("%s Fail to alloc vptr node\n", mmu->name);
+		return 0;
+	}
+
 	iopte = xzalloc_bytes(PAGE_SIZE);
 	if (!iopte) {
 		printk("%s Fail to alloc 2nd level table\n", mmu->name);
@@ -238,10 +251,27 @@ static u32 mmu_iopte_alloc(struct mmu_info *mmu, struct domain *dom, u32 iopgd,
 		iopte[i] = vaddr | IOPTE_SMALL;
 	}
 
+	/* store pointer for following cleanup */
+	alloc_node->vptr = iopte;
+	list_add(&alloc_node->node, &mmu->alloc_list);
+
 	flush_xen_dcache_va_range(iopte, PAGE_SIZE);
 	return __pa(iopte) | IOPGD_TABLE;
 }
 
+static void mmu_cleanup_pagetable(struct mmu_info *mmu)
+{
+	struct mmu_alloc_node *mmu_alloc, *tmp;
+
+	ASSERT(mmu);
+
+	list_for_each_entry_safe(mmu_alloc, tmp, &mmu->alloc_list, node) {
+		xfree(mmu_alloc->vptr);
+		list_del(&mmu_alloc->node);
+		xfree(mmu_alloc);
+	}
+}
+
 /*
  * on boot table is empty
  */
@@ -254,6 +284,9 @@ static int mmu_translate_pagetable(struct domain *dom, struct mmu_info *mmu)
 	ASSERT(dom);
 	ASSERT(mmu);
 
+	/* free all previous allocations */
+	mmu_cleanup_pagetable(mmu);
+
 	/* copy pagetable from  domain to xen */
 	res = mmu_copy_pagetable(mmu);
 	if (res) {
@@ -432,6 +465,8 @@ static int mmu_init(struct mmu_info *mmu, u32 data)
 	printk("%s: %s private pagetable %lu bytes\n",
 		   __func__, mmu->name, IOPGD_TABLE_SIZE);
 
+	INIT_LIST_HEAD(&mmu->alloc_list);
+
 	return 0;
 }
 
-- 
1.7.9.5

Re: [RFC v01 0/3] xen/arm: introduce IOMMU driver for OMAP platforms

[Stefano Stabellini]

On Wed, 22 Jan 2014, Andrii Tseglytskyi wrote:
> Hi,
> 
> The following patch series is an RFC for possible implementation of simple MMU module,
> which is designed to translate IPA to MA for peripheral processors like GPU / IPU
> for OMAP platforms. Currently on our OMAP platform (OMAP5 panda) we have 3 external MMUs
> which need to be handled properly.
> 
> It would be great to get a community feedback - will this be useful for Xen project?
> 
> Let me describe an algorithm briefly. It is simple and straightforward.
> The following simple logic is used to translate addresses from IPA to MA:
> 
> 1. During boot time guest domain creates "pagetable" for external MMU IP.
> Pagetable is a singletone data structure, which is stored in ususal kernel
> heap memory. All memory mappings for corresponding MMU are stored inside it.
> Format of "pagetable" is well defined.
> 
> 2. Guest domain enables peripheral remote processor. As a part of enable sequence
> kernel allocates chunks of heap memory needed for remote processor and stores
> pointers to allocated chunks in already created "pagetable". After it writes
> a physical address of pagetable to MMU configuration register. As result MMU IP
> knows about all allocations, and remote processor can use them directly in its
> software.
> 
> 3. Xen omap mmu driver creates a trap for access to MMU configuration registers.
> It reads a physical address of "pagetable" from MMU register and creates a copy
> of it in own memory. As result - we have two similar configuration data structures -
> first - in guest domain kernel, second - in Xen hypervisor.
> 
> 4. Xen omap mmu driver parses its own copy of pagetable and translate all physical
> addresses to corresponding machine addresses using existing p2m API call.
> After it writes a physical address  of its pagetable (with already translated PA to MA)
> to MMU IP configuration registers and returns control to guest domain.
> 
> As a result - guest domain continues enabling remote processor with it MMU and MMU
> will use new pagetable, modified by Xen omap mmu driver. New pagetable will be used
> directly by MMU IP, and its new structure will be hidden for guest domain kernel,
> it won't know anything about p2m translation.

Why don't you map Dom0 1:1 instead?
If you enabled PLATFORM_QUIRK_DOM0_MAPPING_11 (now enabled by default on
all platforms), all this wouldn't be necessary, right?



> Verified with Xen 4.4-unstable, Linux kernel 3.8 as Dom0, Linux(Android) kernel 3.4 as DomU.
> Target platform OMAP5 panda.
> 
> Thank you for your attention,
> 
> Regards,
> 
> Andrii Tseglytskyi (3):
>   arm: omap: introduce iommu module
>   arm: omap: translate iommu mapping to 4K pages
>   arm: omap: cleanup iopte allocations
> 
>  xen/arch/arm/Makefile     |    1 +
>  xen/arch/arm/io.c         |    1 +
>  xen/arch/arm/io.h         |    1 +
>  xen/arch/arm/omap_iommu.c |  492 +++++++++++++++++++++++++++++++++++++++++++++
>  4 files changed, 495 insertions(+)
>  create mode 100644 xen/arch/arm/omap_iommu.c
> 
> -- 
> 1.7.9.5
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel
> 

Re: [RFC v01 0/3] xen/arm: introduce IOMMU driver for OMAP platforms

[Stefano Stabellini]

On Wed, 22 Jan 2014, Stefano Stabellini wrote:
> On Wed, 22 Jan 2014, Andrii Tseglytskyi wrote:
> > Hi,
> > 
> > The following patch series is an RFC for possible implementation of simple MMU module,
> > which is designed to translate IPA to MA for peripheral processors like GPU / IPU
> > for OMAP platforms. Currently on our OMAP platform (OMAP5 panda) we have 3 external MMUs
> > which need to be handled properly.
> > 
> > It would be great to get a community feedback - will this be useful for Xen project?
> > 
> > Let me describe an algorithm briefly. It is simple and straightforward.
> > The following simple logic is used to translate addresses from IPA to MA:
> > 
> > 1. During boot time guest domain creates "pagetable" for external MMU IP.
> > Pagetable is a singletone data structure, which is stored in ususal kernel
> > heap memory. All memory mappings for corresponding MMU are stored inside it.
> > Format of "pagetable" is well defined.
> > 
> > 2. Guest domain enables peripheral remote processor. As a part of enable sequence
> > kernel allocates chunks of heap memory needed for remote processor and stores
> > pointers to allocated chunks in already created "pagetable". After it writes
> > a physical address of pagetable to MMU configuration register. As result MMU IP
> > knows about all allocations, and remote processor can use them directly in its
> > software.
> > 
> > 3. Xen omap mmu driver creates a trap for access to MMU configuration registers.
> > It reads a physical address of "pagetable" from MMU register and creates a copy
> > of it in own memory. As result - we have two similar configuration data structures -
> > first - in guest domain kernel, second - in Xen hypervisor.
> > 
> > 4. Xen omap mmu driver parses its own copy of pagetable and translate all physical
> > addresses to corresponding machine addresses using existing p2m API call.
> > After it writes a physical address  of its pagetable (with already translated PA to MA)
> > to MMU IP configuration registers and returns control to guest domain.
> > 
> > As a result - guest domain continues enabling remote processor with it MMU and MMU
> > will use new pagetable, modified by Xen omap mmu driver. New pagetable will be used
> > directly by MMU IP, and its new structure will be hidden for guest domain kernel,
> > it won't know anything about p2m translation.
> 
> Why don't you map Dom0 1:1 instead?
> If you enabled PLATFORM_QUIRK_DOM0_MAPPING_11 (now enabled by default on
> all platforms), all this wouldn't be necessary, right?

I guess you can't do just use the 1:1 because you are assigning the GPU
or IPU to a guest other than Dom0, right?


> 
> 
> > Verified with Xen 4.4-unstable, Linux kernel 3.8 as Dom0, Linux(Android) kernel 3.4 as DomU.
> > Target platform OMAP5 panda.
> > 
> > Thank you for your attention,
> > 
> > Regards,
> > 
> > Andrii Tseglytskyi (3):
> >   arm: omap: introduce iommu module
> >   arm: omap: translate iommu mapping to 4K pages
> >   arm: omap: cleanup iopte allocations
> > 
> >  xen/arch/arm/Makefile     |    1 +
> >  xen/arch/arm/io.c         |    1 +
> >  xen/arch/arm/io.h         |    1 +
> >  xen/arch/arm/omap_iommu.c |  492 +++++++++++++++++++++++++++++++++++++++++++++
> >  4 files changed, 495 insertions(+)
> >  create mode 100644 xen/arch/arm/omap_iommu.c
> > 
> > -- 
> > 1.7.9.5
> > 
> > 
> > _______________________________________________
> > Xen-devel mailing list
> > Xen-devel@lists.xen.org
> > http://lists.xen.org/xen-devel
> > 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel
> 

Re: [RFC v01 0/3] xen/arm: introduce IOMMU driver for OMAP platforms

[Andrii Tseglytskyi]

[-- Attachment #1.1: Type: text/plain, Size: 324 bytes --]

Hi Stefano,


>
> I guess you can't do just use the 1:1 because you are assigning the GPU
> or IPU to a guest other than Dom0, right?
>
>
Right.  GPU / IPU are both assigned to DomU, which is Android 4.3.

Thank you for your answer,

regards,
Andrii



-- 

Andrii Tseglytskyi | Embedded Dev
GlobalLogic
www.globallogic.com

[-- Attachment #1.2: Type: text/html, Size: 1559 bytes --]

[-- Attachment #2: Type: text/plain, Size: 126 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Re: [RFC v01 0/3] xen/arm: introduce IOMMU driver for OMAP platforms

[Andrii Tseglytskyi]

Hi Stefano,

On Wed, Jan 22, 2014 at 7:39 PM, Stefano Stabellini
<stefano.stabellini@eu.citrix.com> wrote:
> On Wed, 22 Jan 2014, Stefano Stabellini wrote:
>
> I guess you can't do just use the 1:1 because you are assigning the GPU
> or IPU to a guest other than Dom0, right?
>

Right.  GPU / IPU are both assigned to DomU, which is Android 4.3.

Thank you for your answer,

regards,
Andrii


-- 

Andrii Tseglytskyi | Embedded Dev
GlobalLogic
www.globallogic.com