[PATCH v5 0/4] Renesas IPMMU driver for sh7372, sh73a0 and r8a7740

[PATCH v5 0/4] Renesas IPMMU driver for sh7372, sh73a0 and r8a7740

[Hideki EIRAKU]

This is the Renesas IPMMU driver, IOMMU API implementation and IPMMU
device support for sh7372 (AP4EVB and Mackerel), sh73a0 and r8a7740.

The IPMMU module supports the MMU function and the PMB function.  The
MMU function provides address translation by pagetable compatible with
ARMv6.  The PMB function provides address translation including
tile-linear translation.  This is implementation of the MMU function.

Note: The IPMMU module translates DMA memory address from all of the
IP blocks on the ICB.  The IOMMU API has iommu_attach_device() which
attaches one device to one domain, but the IPMMU module cannot attach
only one device.  The IPMMU module only can attach all devices to one
domain.

Tested on:
- armadillo800eva
- kzma9
- mackerel (I could not make it boot the v3.8-rc1 with Device Tree, so
  I tested a kernel without Device Tree support, reverting
  3b7b70552afe351a8bd8fff1eb2d60aab2206576 and
  0ce53cdc5c7e28f378e480363a0b0c2ed7e7eaf9)

How to test:
1. Boot a kernel built with CONFIG_SHMOBILE_IOMMU=y.
2. Check whether a display on the board shows console normally.
3. Check whether the IMCTR1 register is 1 which means the MMU function enabled.
   To read the IMCTR1 register:
   # peekpoke 0xfe951000
4. Change the IMCTR1 to 0 to disable the MMU function.
   To change the IMCTR1 register to 0:
   # peekpoke 0xfe951000 0
5. Check whether the display shows broken data.
6. Change the IMCTR1 to 1 to enable the MMU function.
   To change the IMCTR1 register to 1:
   # peekpoke 0xfe951000 1
7. Check whether a display on the board shows console normally.
8. Capture and check the framebuffer contents via memory mapped by mmap().
   To capture the framebuffer contents and save the image to a file "filename":
   # fbcat /dev/fb0 > filename
   To look at the contents saved by fbcat on a PC:
   $ display filename
- fbcat can be downloaded from https://code.google.com/p/fbcat/.
- display is one of the ImageMagick tools.
- peekpoke is:
#include <sys/mman.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdlib.h>
#include <stdint.h>
#include <stdio.h>
#include <err.h>
int main(int argc,char**argv){
  uint32_t*p,addr;
  int fd=open("/dev/mem",O_RDWR);
  if(fd<0)err(1,"open");
  addr=strtoul(argv[1],NULL,0);
  p=mmap(NULL,0x1000,PROT_READ|PROT_WRITE,MAP_SHARED,fd,addr&~0xfff);
  if(p==MAP_FAILED)err(1,"mmap");
  if(argc<=2)printf("0x%08x\n",p[addr/4&0x3ff]);
  else printf("wrote 0x%08x\n",p[addr/4&0x3ff]=strtoul(argv[2],NULL,0));
}

Changelog:
v5:
- Rebased on 3.8-rc1.
- Add sh73a0 and r8a7740 support.
- CONFIG_SHMOBILE_IOMMU is now selectable only for arm shmobile
  architecture.
- Fix "defined but not used" warnings when CONFIG_SHMOBILE_IOMMU is
  not set.
- Add help text of CONFIG_SHMOBILE_IOMMU.
- Apply some of Laurent Pinchart's rework patches: Move the driver to
  drivers/iommu/.
- Use platform_data to specify attaching devices instead of
  ipmmu_add_device() API.
v4:
- Rebased on 3.7-rc1.
- Use address space size instead of page table size in config.
- Use Kconfig default value instead of #ifdef-#define-#endif.
v3:
- Rebased on 3.6-rc5.
- Simplify configs.  SHMOBILE_IPMMU is now selected by setting
  SHMOBILE_IOMMU.
- Remove weak symbols.
- Use drvdata to store private driver data.
- Make a function for writing to a register of IPMMU.
- Add a lock to accessing the tlb_enabled member.
- Make unmap work correctly with size larger than map size.
- Free L2 page table when 1MiB page is mapped or unmapped.
- Add VEU devices as IP blocks on the ICB.
v2:
- Rebased on v3.6-rc1.
- Make variable names clear.
- Page table size can now be selected by config.

Hideki EIRAKU (4):
  iommu/shmobile: Add iommu driver for Renesas IPMMU modules
  ARM: mach-shmobile: sh7372: Add IPMMU device
  ARM: mach-shmobile: sh73a0: Add IPMMU device
  ARM: mach-shmobile: r8a7740: Add IPMMU device

 arch/arm/mach-shmobile/setup-r8a7740.c |  33 ++++
 arch/arm/mach-shmobile/setup-sh7372.c  |  39 ++++
 arch/arm/mach-shmobile/setup-sh73a0.c  |  31 +++
 drivers/iommu/Kconfig                  |  74 +++++++
 drivers/iommu/Makefile                 |   2 +
 drivers/iommu/shmobile-iommu.c         | 341 +++++++++++++++++++++++++++++++++
 drivers/iommu/shmobile-ipmmu.c         | 136 +++++++++++++
 drivers/iommu/shmobile-ipmmu.h         |  37 ++++
 include/linux/platform_data/sh_ipmmu.h |  18 ++
 9 files changed, 711 insertions(+)
 create mode 100644 drivers/iommu/shmobile-iommu.c
 create mode 100644 drivers/iommu/shmobile-ipmmu.c
 create mode 100644 drivers/iommu/shmobile-ipmmu.h
 create mode 100644 include/linux/platform_data/sh_ipmmu.h

-- 
1.8.0

[PATCH v5 1/4] iommu/shmobile: Add iommu driver for Renesas IPMMU modules

[Hideki EIRAKU]

This is the Renesas IPMMU driver and IOMMU API implementation.

The IPMMU module supports the MMU function and the PMB function.  The
MMU function provides address translation by pagetable compatible with
ARMv6.  The PMB function provides address translation including
tile-linear translation.  This patch implements the MMU function.

The iommu driver does not register a platform driver directly because:
- the register space of the MMU function and the PMB function
  have a common register (used for settings flush), so they should ideally
  have a way to appropriately share this register.
- the MMU function uses the IOMMU API while the PMB function does not.
- the two functions may be used independently.

Signed-off-by: Hideki EIRAKU <hdk@igel.co.jp>
---
 drivers/iommu/Kconfig                  |  74 +++++++
 drivers/iommu/Makefile                 |   2 +
 drivers/iommu/shmobile-iommu.c         | 341 +++++++++++++++++++++++++++++++++
 drivers/iommu/shmobile-ipmmu.c         | 136 +++++++++++++
 drivers/iommu/shmobile-ipmmu.h         |  37 ++++
 include/linux/platform_data/sh_ipmmu.h |  18 ++
 6 files changed, 608 insertions(+)
 create mode 100644 drivers/iommu/shmobile-iommu.c
 create mode 100644 drivers/iommu/shmobile-ipmmu.c
 create mode 100644 drivers/iommu/shmobile-ipmmu.h
 create mode 100644 include/linux/platform_data/sh_ipmmu.h

diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
index e39f9db..d364494 100644
--- a/drivers/iommu/Kconfig
+++ b/drivers/iommu/Kconfig
@@ -187,4 +187,78 @@ config EXYNOS_IOMMU_DEBUG
 
 	  Say N unless you need kernel log message for IOMMU debugging
 
+config SHMOBILE_IPMMU
+	bool
+
+config SHMOBILE_IPMMU_TLB
+	bool
+
+config SHMOBILE_IOMMU
+	bool "IOMMU for Renesas IPMMU/IPMMUI"
+	default n
+	depends on (ARM && ARCH_SHMOBILE)
+	select IOMMU_API
+	select ARM_DMA_USE_IOMMU
+	select SHMOBILE_IPMMU
+	select SHMOBILE_IPMMU_TLB
+	help
+	  Support for Renesas IPMMU/IPMMUI. This option enables
+	  remapping of DMA memory accesses from all of the IP blocks
+	  on the ICB.
+
+	  Warning: Drivers (including userspace drivers of UIO
+	  devices) of the IP blocks on the ICB *must* use addresses
+	  allocated from the IPMMU (iova) for DMA with this option
+	  enabled.
+
+	  If unsure, say N.
+
+choice
+	prompt "IPMMU/IPMMUI address space size"
+	default SHMOBILE_IOMMU_ADDRSIZE_2048MB
+	depends on SHMOBILE_IOMMU
+	help
+	  This option sets IPMMU/IPMMUI address space size by
+	  adjusting the 1st level page table size. The page table size
+	  is calculated as follows:
+
+	      page table size = number of page table entries * 4 bytes
+	      number of page table entries = address space size / 1 MiB
+
+	  For example, when the address space size is 2048 MiB, the
+	  1st level page table size is 8192 bytes.
+
+	config SHMOBILE_IOMMU_ADDRSIZE_2048MB
+		bool "2 GiB"
+
+	config SHMOBILE_IOMMU_ADDRSIZE_1024MB
+		bool "1 GiB"
+
+	config SHMOBILE_IOMMU_ADDRSIZE_512MB
+		bool "512 MiB"
+
+	config SHMOBILE_IOMMU_ADDRSIZE_256MB
+		bool "256 MiB"
+
+	config SHMOBILE_IOMMU_ADDRSIZE_128MB
+		bool "128 MiB"
+
+	config SHMOBILE_IOMMU_ADDRSIZE_64MB
+		bool "64 MiB"
+
+	config SHMOBILE_IOMMU_ADDRSIZE_32MB
+		bool "32 MiB"
+
+endchoice
+
+config SHMOBILE_IOMMU_L1SIZE
+	int
+	default 8192 if SHMOBILE_IOMMU_ADDRSIZE_2048MB
+	default 4096 if SHMOBILE_IOMMU_ADDRSIZE_1024MB
+	default 2048 if SHMOBILE_IOMMU_ADDRSIZE_512MB
+	default 1024 if SHMOBILE_IOMMU_ADDRSIZE_256MB
+	default 512 if SHMOBILE_IOMMU_ADDRSIZE_128MB
+	default 256 if SHMOBILE_IOMMU_ADDRSIZE_64MB
+	default 128 if SHMOBILE_IOMMU_ADDRSIZE_32MB
+
 endif # IOMMU_SUPPORT
diff --git a/drivers/iommu/Makefile b/drivers/iommu/Makefile
index f66b816..ef0e520 100644
--- a/drivers/iommu/Makefile
+++ b/drivers/iommu/Makefile
@@ -13,3 +13,5 @@ obj-$(CONFIG_OMAP_IOMMU_DEBUG) += omap-iommu-debug.o
 obj-$(CONFIG_TEGRA_IOMMU_GART) += tegra-gart.o
 obj-$(CONFIG_TEGRA_IOMMU_SMMU) += tegra-smmu.o
 obj-$(CONFIG_EXYNOS_IOMMU) += exynos-iommu.o
+obj-$(CONFIG_SHMOBILE_IOMMU) += shmobile-iommu.o
+obj-$(CONFIG_SHMOBILE_IPMMU) += shmobile-ipmmu.o
diff --git a/drivers/iommu/shmobile-iommu.c b/drivers/iommu/shmobile-iommu.c
new file mode 100644
index 0000000..360a06f
--- /dev/null
+++ b/drivers/iommu/shmobile-iommu.c
@@ -0,0 +1,341 @@
+/*
+ * IOMMU for IPMMU/IPMMUI
+ * Copyright (C) 2012  Hideki EIRAKU
+ *
+ * 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; version 2 of the License.
+ */
+
+#include <linux/dma-mapping.h>
+#include <linux/dmapool.h>
+#include <linux/io.h>
+#include <linux/iommu.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+#include <asm/dma-iommu.h>
+#include "shmobile-ipmmu.h"
+
+#define L1_SIZE CONFIG_SHMOBILE_IOMMU_L1SIZE
+#define L1_LEN (L1_SIZE / 4)
+#define L1_ALIGN L1_SIZE
+#define L2_SIZE 0x400
+#define L2_LEN (L2_SIZE / 4)
+#define L2_ALIGN L2_SIZE
+
+struct shmobile_iommu_domain_pgtable {
+	uint32_t *pgtable;
+	dma_addr_t handle;
+};
+
+struct shmobile_iommu_domain {
+	struct shmobile_ipmmu *ipmmu;
+	struct shmobile_iommu_domain_pgtable l1, l2[L1_LEN];
+	spinlock_t map_lock;
+	atomic_t active;
+};
+
+static struct shmobile_ipmmu *platform_bus_ipmmu;
+static struct dma_pool *l1pool, *l2pool;
+static spinlock_t lock;
+static struct shmobile_iommu_domain *attached;
+static int num_attached_devices;
+
+static int shmobile_iommu_domain_init(struct iommu_domain *domain)
+{
+	struct shmobile_iommu_domain *sh_domain;
+	int i;
+
+	sh_domain = kmalloc(sizeof(*sh_domain), GFP_KERNEL);
+	if (!sh_domain)
+		return -ENOMEM;
+	sh_domain->l1.pgtable = dma_pool_alloc(l1pool, GFP_KERNEL,
+					       &sh_domain->l1.handle);
+	if (!sh_domain->l1.pgtable) {
+		kfree(sh_domain);
+		return -ENOMEM;
+	}
+	for (i = 0; i < L1_LEN; i++)
+		sh_domain->l2[i].pgtable = NULL;
+	memset(sh_domain->l1.pgtable, 0, L1_SIZE);
+	spin_lock_init(&sh_domain->map_lock);
+	atomic_set(&sh_domain->active, 0);
+	domain->priv = sh_domain;
+	return 0;
+}
+
+static void shmobile_iommu_domain_destroy(struct iommu_domain *domain)
+{
+	struct shmobile_iommu_domain *sh_domain = domain->priv;
+	int i;
+
+	for (i = 0; i < L1_LEN; i++) {
+		if (sh_domain->l2[i].pgtable)
+			dma_pool_free(l2pool, sh_domain->l2[i].pgtable,
+				      sh_domain->l2[i].handle);
+	}
+	dma_pool_free(l1pool, sh_domain->l1.pgtable, sh_domain->l1.handle);
+	kfree(sh_domain);
+	domain->priv = NULL;
+}
+
+static int shmobile_iommu_attach_device(struct iommu_domain *domain,
+					struct device *dev)
+{
+	struct shmobile_iommu_domain *sh_domain = domain->priv;
+	struct shmobile_ipmmu *ipmmu = platform_bus_ipmmu;
+	int ret = -EBUSY;
+
+	spin_lock(&lock);
+	if (attached != sh_domain) {
+		if (attached)
+			goto err;
+		atomic_set(&sh_domain->active, 1);
+		ipmmu_tlb_set(ipmmu, sh_domain->l1.handle, L1_SIZE, 0);
+		wmb();
+		ipmmu_tlb_flush(ipmmu);
+		attached = sh_domain;
+		num_attached_devices = 0;
+		sh_domain->ipmmu = ipmmu;
+	}
+	num_attached_devices++;
+	ret = 0;
+err:
+	spin_unlock(&lock);
+	return ret;
+}
+
+static void shmobile_iommu_detach_device(struct iommu_domain *domain,
+					 struct device *dev)
+{
+	struct shmobile_iommu_domain *sh_domain = domain->priv;
+	struct shmobile_ipmmu *ipmmu = platform_bus_ipmmu;
+
+	spin_lock(&lock);
+	atomic_set(&sh_domain->active, 0);
+	num_attached_devices--;
+	if (!num_attached_devices) {
+		ipmmu_tlb_set(ipmmu, 0, 0, 0);
+		ipmmu_tlb_flush(ipmmu);
+		attached = NULL;
+		sh_domain->ipmmu = NULL;
+	}
+	spin_unlock(&lock);
+}
+
+static int
+l2alloc(struct shmobile_iommu_domain *sh_domain, unsigned int l1index)
+{
+	if (!sh_domain->l2[l1index].pgtable) {
+		sh_domain->l2[l1index].pgtable = dma_pool_alloc(l2pool,
+						GFP_KERNEL,
+						&sh_domain->l2[l1index].handle);
+		if (!sh_domain->l2[l1index].pgtable)
+			return -ENOMEM;
+		memset(sh_domain->l2[l1index].pgtable, 0, L2_SIZE);
+	}
+	sh_domain->l1.pgtable[l1index] = sh_domain->l2[l1index].handle | 0x1;
+	return 0;
+}
+
+static void
+l2realfree(struct shmobile_iommu_domain_pgtable *l2)
+{
+	if (l2->pgtable)
+		dma_pool_free(l2pool, l2->pgtable, l2->handle);
+}
+
+static int
+l2free(struct shmobile_iommu_domain *sh_domain, unsigned int l1index,
+	struct shmobile_iommu_domain_pgtable *l2)
+{
+	sh_domain->l1.pgtable[l1index] = 0;
+	if (sh_domain->l2[l1index].pgtable) {
+		*l2 = sh_domain->l2[l1index];
+		sh_domain->l2[l1index].pgtable = NULL;
+	}
+	return 0;
+}
+
+static int shmobile_iommu_map(struct iommu_domain *domain, unsigned long iova,
+			      phys_addr_t paddr, size_t size, int prot)
+{
+	struct shmobile_iommu_domain_pgtable l2 = { .pgtable = NULL };
+	struct shmobile_iommu_domain *sh_domain = domain->priv;
+	unsigned int l1index, l2index, i;
+	int ret;
+
+	l1index = iova >> 20;
+	switch (size) {
+	case 0x1000:
+		l2index = (iova >> 12) & 0xff;
+		spin_lock(&sh_domain->map_lock);
+		ret = l2alloc(sh_domain, l1index);
+		if (!ret)
+			sh_domain->l2[l1index].pgtable[l2index] = paddr | 0xff2;
+		spin_unlock(&sh_domain->map_lock);
+		break;
+	case 0x10000:
+		l2index = (iova >> 12) & 0xf0;
+		spin_lock(&sh_domain->map_lock);
+		ret = l2alloc(sh_domain, l1index);
+		if (!ret) {
+			for (i = 0; i < 0x10; i++)
+				sh_domain->l2[l1index].pgtable[l2index + i] =
+					paddr | 0xff1;
+		}
+		spin_unlock(&sh_domain->map_lock);
+		break;
+	case 0x100000:
+		spin_lock(&sh_domain->map_lock);
+		l2free(sh_domain, l1index, &l2);
+		sh_domain->l1.pgtable[l1index] = paddr | 0xc02;
+		spin_unlock(&sh_domain->map_lock);
+		ret = 0;
+		break;
+	default:
+		ret = -EINVAL;
+	}
+	if (!ret && atomic_read(&sh_domain->active)) {
+		wmb();
+		ipmmu_tlb_flush(sh_domain->ipmmu);
+		l2realfree(&l2);
+	}
+	return ret;
+}
+
+static size_t shmobile_iommu_unmap(struct iommu_domain *domain,
+				   unsigned long iova, size_t size)
+{
+	struct shmobile_iommu_domain_pgtable l2 = { .pgtable = NULL };
+	struct shmobile_iommu_domain *sh_domain = domain->priv;
+	unsigned int l1index, l2index, i;
+	uint32_t l2entry = 0;
+	size_t ret = 0;
+
+	l1index = iova >> 20;
+	if (!(iova & 0xFFFFF) && size >= 0x100000) {
+		spin_lock(&sh_domain->map_lock);
+		l2free(sh_domain, l1index, &l2);
+		spin_unlock(&sh_domain->map_lock);
+		ret = 0x100000;
+		goto done;
+	}
+	l2index = (iova >> 12) & 0xff;
+	spin_lock(&sh_domain->map_lock);
+	if (sh_domain->l2[l1index].pgtable)
+		l2entry = sh_domain->l2[l1index].pgtable[l2index];
+	switch (l2entry & 3) {
+	case 1:
+		if (l2index & 0xf)
+			break;
+		for (i = 0; i < 0x10; i++)
+			sh_domain->l2[l1index].pgtable[l2index + i] = 0;
+		ret = 0x10000;
+		break;
+	case 2:
+		sh_domain->l2[l1index].pgtable[l2index] = 0;
+		ret = 0x1000;
+		break;
+	}
+	spin_unlock(&sh_domain->map_lock);
+done:
+	if (ret && atomic_read(&sh_domain->active)) {
+		wmb();
+		ipmmu_tlb_flush(sh_domain->ipmmu);
+		l2realfree(&l2);
+	}
+	return ret;
+}
+
+static phys_addr_t shmobile_iommu_iova_to_phys(struct iommu_domain *domain,
+					       unsigned long iova)
+{
+	struct shmobile_iommu_domain *sh_domain = domain->priv;
+	uint32_t l1entry = 0, l2entry = 0;
+	unsigned int l1index, l2index;
+
+	l1index = iova >> 20;
+	l2index = (iova >> 12) & 0xff;
+	spin_lock(&sh_domain->map_lock);
+	if (sh_domain->l2[l1index].pgtable)
+		l2entry = sh_domain->l2[l1index].pgtable[l2index];
+	else
+		l1entry = sh_domain->l1.pgtable[l1index];
+	spin_unlock(&sh_domain->map_lock);
+	switch (l2entry & 3) {
+	case 1:
+		return (l2entry & ~0xffff) | (iova & 0xffff);
+	case 2:
+		return (l2entry & ~0xfff) | (iova & 0xfff);
+	default:
+		if ((l1entry & 3) == 2)
+			return (l1entry & ~0xfffff) | (iova & 0xfffff);
+		return 0;
+	}
+}
+
+static int find_dev_name(struct shmobile_ipmmu *ipmmu, const char *dev_name)
+{
+	unsigned int i, n = ipmmu->num_dev_names;
+
+	for (i = 0; i < n; i++) {
+		if (strcmp(ipmmu->dev_names[i], dev_name) == 0)
+			return 1;
+	}
+	return 0;
+}
+
+static int shmobile_iommu_add_device(struct device *dev)
+{
+	struct shmobile_ipmmu *ipmmu = platform_bus_ipmmu;
+	struct dma_iommu_mapping *mapping;
+
+	if (!find_dev_name(ipmmu, dev_name(dev)))
+		return 0;
+	mapping = ipmmu->iommu_mapping;
+	if (!mapping) {
+		mapping = arm_iommu_create_mapping(&platform_bus_type, 0,
+						   L1_LEN << 20, 0);
+		if (IS_ERR(mapping))
+			return PTR_ERR(mapping);
+		ipmmu->iommu_mapping = mapping;
+	}
+	if (arm_iommu_attach_device(dev, mapping))
+		pr_err("arm_iommu_attach_device failed\n");
+	return 0;
+}
+
+static struct iommu_ops shmobile_iommu_ops = {
+	.domain_init = shmobile_iommu_domain_init,
+	.domain_destroy = shmobile_iommu_domain_destroy,
+	.attach_dev = shmobile_iommu_attach_device,
+	.detach_dev = shmobile_iommu_detach_device,
+	.map = shmobile_iommu_map,
+	.unmap = shmobile_iommu_unmap,
+	.iova_to_phys = shmobile_iommu_iova_to_phys,
+	.add_device = shmobile_iommu_add_device,
+	.pgsize_bitmap = 0x111000,
+};
+
+int ipmmu_iommu_init(struct shmobile_ipmmu *ipmmu)
+{
+	dma_set_coherent_mask(ipmmu->dev, DMA_BIT_MASK(32));
+	l1pool = dma_pool_create("shmobile-iommu-pgtable1", ipmmu->dev,
+				 L1_SIZE, L1_ALIGN, 0);
+	if (!l1pool)
+		goto nomem_pool1;
+	l2pool = dma_pool_create("shmobile-iommu-pgtable2", ipmmu->dev,
+				 L2_SIZE, L2_ALIGN, 0);
+	if (!l2pool)
+		goto nomem_pool2;
+	spin_lock_init(&lock);
+	attached = NULL;
+	platform_bus_ipmmu = ipmmu;
+	bus_set_iommu(&platform_bus_type, &shmobile_iommu_ops);
+	return 0;
+nomem_pool2:
+	dma_pool_destroy(l1pool);
+nomem_pool1:
+	return -ENOMEM;
+}
diff --git a/drivers/iommu/shmobile-ipmmu.c b/drivers/iommu/shmobile-ipmmu.c
new file mode 100644
index 0000000..8321f89
--- /dev/null
+++ b/drivers/iommu/shmobile-ipmmu.c
@@ -0,0 +1,136 @@
+/*
+ * IPMMU/IPMMUI
+ * Copyright (C) 2012  Hideki EIRAKU
+ *
+ * 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; version 2 of the License.
+ */
+
+#include <linux/err.h>
+#include <linux/export.h>
+#include <linux/io.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+#include <linux/platform_data/sh_ipmmu.h>
+#include "shmobile-ipmmu.h"
+
+#define IMCTR1 0x000
+#define IMCTR2 0x004
+#define IMASID 0x010
+#define IMTTBR 0x014
+#define IMTTBCR 0x018
+
+#define IMCTR1_TLBEN (1 << 0)
+#define IMCTR1_FLUSH (1 << 1)
+
+static void ipmmu_reg_write(struct shmobile_ipmmu *ipmmu, unsigned long reg_off,
+			    unsigned long data)
+{
+	iowrite32(data, ipmmu->ipmmu_base + reg_off);
+}
+
+void ipmmu_tlb_flush(struct shmobile_ipmmu *ipmmu)
+{
+	if (!ipmmu)
+		return;
+
+	mutex_lock(&ipmmu->flush_lock);
+	if (ipmmu->tlb_enabled)
+		ipmmu_reg_write(ipmmu, IMCTR1, IMCTR1_FLUSH | IMCTR1_TLBEN);
+	else
+		ipmmu_reg_write(ipmmu, IMCTR1, IMCTR1_FLUSH);
+	mutex_unlock(&ipmmu->flush_lock);
+}
+
+void ipmmu_tlb_set(struct shmobile_ipmmu *ipmmu, unsigned long phys, int size,
+		   int asid)
+{
+	if (!ipmmu)
+		return;
+
+	mutex_lock(&ipmmu->flush_lock);
+	switch (size) {
+	default:
+		ipmmu->tlb_enabled = 0;
+		break;
+	case 0x2000:
+		ipmmu_reg_write(ipmmu, IMTTBCR, 1);
+		ipmmu->tlb_enabled = 1;
+		break;
+	case 0x1000:
+		ipmmu_reg_write(ipmmu, IMTTBCR, 2);
+		ipmmu->tlb_enabled = 1;
+		break;
+	case 0x800:
+		ipmmu_reg_write(ipmmu, IMTTBCR, 3);
+		ipmmu->tlb_enabled = 1;
+		break;
+	case 0x400:
+		ipmmu_reg_write(ipmmu, IMTTBCR, 4);
+		ipmmu->tlb_enabled = 1;
+		break;
+	case 0x200:
+		ipmmu_reg_write(ipmmu, IMTTBCR, 5);
+		ipmmu->tlb_enabled = 1;
+		break;
+	case 0x100:
+		ipmmu_reg_write(ipmmu, IMTTBCR, 6);
+		ipmmu->tlb_enabled = 1;
+		break;
+	case 0x80:
+		ipmmu_reg_write(ipmmu, IMTTBCR, 7);
+		ipmmu->tlb_enabled = 1;
+		break;
+	}
+	ipmmu_reg_write(ipmmu, IMTTBR, phys);
+	ipmmu_reg_write(ipmmu, IMASID, asid);
+	mutex_unlock(&ipmmu->flush_lock);
+}
+
+static int ipmmu_probe(struct platform_device *pdev)
+{
+	struct shmobile_ipmmu *ipmmu;
+	struct resource *res;
+	struct shmobile_ipmmu_platform_data *pdata = pdev->dev.platform_data;
+
+	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	if (!res) {
+		dev_err(&pdev->dev, "cannot get platform resources\n");
+		return -ENOENT;
+	}
+	ipmmu = devm_kzalloc(&pdev->dev, sizeof(*ipmmu), GFP_KERNEL);
+	if (!ipmmu) {
+		dev_err(&pdev->dev, "cannot allocate device data\n");
+		return -ENOMEM;
+	}
+	mutex_init(&ipmmu->flush_lock);
+	ipmmu->dev = &pdev->dev;
+	ipmmu->ipmmu_base = devm_ioremap_nocache(&pdev->dev, res->start,
+						resource_size(res));
+	if (!ipmmu->ipmmu_base) {
+		dev_err(&pdev->dev, "ioremap_nocache failed\n");
+		return -ENOMEM;
+	}
+	ipmmu->dev_names = pdata->dev_names;
+	ipmmu->num_dev_names = pdata->num_dev_names;
+	platform_set_drvdata(pdev, ipmmu);
+	ipmmu_reg_write(ipmmu, IMCTR1, 0x0); /* disable TLB */
+	ipmmu_reg_write(ipmmu, IMCTR2, 0x0); /* disable PMB */
+	ipmmu_iommu_init(ipmmu);
+	return 0;
+}
+
+static struct platform_driver ipmmu_driver = {
+	.probe = ipmmu_probe,
+	.driver = {
+		.owner = THIS_MODULE,
+		.name = "ipmmu",
+	},
+};
+
+static int __init ipmmu_init(void)
+{
+	return platform_driver_register(&ipmmu_driver);
+}
+subsys_initcall(ipmmu_init);
diff --git a/drivers/iommu/shmobile-ipmmu.h b/drivers/iommu/shmobile-ipmmu.h
new file mode 100644
index 0000000..6270e7c
--- /dev/null
+++ b/drivers/iommu/shmobile-ipmmu.h
@@ -0,0 +1,37 @@
+/* shmobile-ipmmu.h
+ *
+ * Copyright (C) 2012  Hideki EIRAKU
+ *
+ * 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; version 2 of the License.
+ */
+
+#ifndef __SHMOBILE_IPMMU_H__
+#define __SHMOBILE_IPMMU_H__
+
+struct dma_iommu_mapping;
+
+struct shmobile_ipmmu {
+	struct device *dev;
+	void __iomem *ipmmu_base;
+	int tlb_enabled;
+	struct mutex flush_lock;
+	struct dma_iommu_mapping *iommu_mapping;
+	const char * const *dev_names;
+	unsigned int num_dev_names;
+};
+
+#ifdef CONFIG_SHMOBILE_IPMMU_TLB
+void ipmmu_tlb_flush(struct shmobile_ipmmu *ipmmu);
+void ipmmu_tlb_set(struct shmobile_ipmmu *ipmmu, unsigned long phys, int size,
+		   int asid);
+int ipmmu_iommu_init(struct shmobile_ipmmu *ipmmu);
+#else
+static int ipmmu_iommu_init(struct shmobile_ipmmu *ipmmu)
+{
+	return -EINVAL;
+}
+#endif
+
+#endif /* __SHMOBILE_IPMMU_H__ */
diff --git a/include/linux/platform_data/sh_ipmmu.h b/include/linux/platform_data/sh_ipmmu.h
new file mode 100644
index 0000000..39f7405
--- /dev/null
+++ b/include/linux/platform_data/sh_ipmmu.h
@@ -0,0 +1,18 @@
+/* sh_ipmmu.h
+ *
+ * Copyright (C) 2012  Hideki EIRAKU
+ *
+ * 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; version 2 of the License.
+ */
+
+#ifndef __SH_IPMMU_H__
+#define __SH_IPMMU_H__
+
+struct shmobile_ipmmu_platform_data {
+	const char * const *dev_names;
+	unsigned int num_dev_names;
+};
+
+#endif /* __SH_IPMMU_H__ */
-- 
1.8.0

[PATCH v5 2/4] ARM: mach-shmobile: sh7372: Add IPMMU device

[Hideki EIRAKU]

This patch adds an IPMMU device and notifies the IPMMU driver which
devices are connected via the IPMMU module.  All devices connected to the main
memory bus via the IPMMU module MUST be registered when SHMOBILE_IPMMU and
SHMOBILE_IOMMU are enabled because physical address cannot be used
while the IPMMU module's MMU function is enabled.

Signed-off-by: Hideki EIRAKU <hdk@igel.co.jp>
---
 arch/arm/mach-shmobile/setup-sh7372.c | 39 +++++++++++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)

diff --git a/arch/arm/mach-shmobile/setup-sh7372.c b/arch/arm/mach-shmobile/setup-sh7372.c
index c917882..04ad98e 100644
--- a/arch/arm/mach-shmobile/setup-sh7372.c
+++ b/arch/arm/mach-shmobile/setup-sh7372.c
@@ -33,6 +33,7 @@
 #include <linux/sh_timer.h>
 #include <linux/pm_domain.h>
 #include <linux/dma-mapping.h>
+#include <linux/platform_data/sh_ipmmu.h>
 #include <mach/dma-register.h>
 #include <mach/hardware.h>
 #include <mach/irqs.h>
@@ -982,6 +983,43 @@ static struct platform_device spu1_device = {
 	.num_resources	= ARRAY_SIZE(spu1_resources),
 };
 
+/* IPMMUI (an IPMMU module for ICB/LMB) */
+static struct resource ipmmu_resources[] = {
+	[0] = {
+		.name	= "IPMMUI",
+		.start	= 0xfe951000,
+		.end	= 0xfe9510ff,
+		.flags	= IORESOURCE_MEM,
+	},
+};
+
+static const char * const ipmmu_dev_names[] = {
+	"sh_mobile_lcdc_fb.0",
+	"sh_mobile_lcdc_fb.1",
+	"sh_mobile_ceu.0",
+	"uio_pdrv_genirq.0",
+	"uio_pdrv_genirq.1",
+	"uio_pdrv_genirq.2",
+	"uio_pdrv_genirq.3",
+	"uio_pdrv_genirq.4",
+	"uio_pdrv_genirq.5",
+};
+
+static struct shmobile_ipmmu_platform_data ipmmu_platform_data = {
+	.dev_names = ipmmu_dev_names,
+	.num_dev_names = ARRAY_SIZE(ipmmu_dev_names),
+};
+
+static struct platform_device ipmmu_device = {
+	.name           = "ipmmu",
+	.id             = -1,
+	.dev = {
+		.platform_data = &ipmmu_platform_data,
+	},
+	.resource       = ipmmu_resources,
+	.num_resources  = ARRAY_SIZE(ipmmu_resources),
+};
+
 static struct platform_device *sh7372_early_devices[] __initdata = {
 	&scif0_device,
 	&scif1_device,
@@ -993,6 +1031,7 @@ static struct platform_device *sh7372_early_devices[] __initdata = {
 	&cmt2_device,
 	&tmu00_device,
 	&tmu01_device,
+	&ipmmu_device,
 };
 
 static struct platform_device *sh7372_late_devices[] __initdata = {
-- 
1.8.0

[PATCH v5 3/4] ARM: mach-shmobile: sh73a0: Add IPMMU device

[Hideki EIRAKU]

This patch adds an IPMMU device and notifies the IPMMU driver which
devices are connected via the IPMMU module.  All devices connected to the main
memory bus via the IPMMU module MUST be registered when SHMOBILE_IPMMU and
SHMOBILE_IOMMU are enabled because physical address cannot be used
while the IPMMU module's MMU function is enabled.

Signed-off-by: Hideki EIRAKU <hdk@igel.co.jp>
---
 arch/arm/mach-shmobile/setup-sh73a0.c | 31 +++++++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)

diff --git a/arch/arm/mach-shmobile/setup-sh73a0.c b/arch/arm/mach-shmobile/setup-sh73a0.c
index db99a4a..36c2b2e 100644
--- a/arch/arm/mach-shmobile/setup-sh73a0.c
+++ b/arch/arm/mach-shmobile/setup-sh73a0.c
@@ -30,6 +30,7 @@
 #include <linux/sh_dma.h>
 #include <linux/sh_intc.h>
 #include <linux/sh_timer.h>
+#include <linux/platform_data/sh_ipmmu.h>
 #include <mach/dma-register.h>
 #include <mach/hardware.h>
 #include <mach/irqs.h>
@@ -754,6 +755,35 @@ static struct platform_device pmu_device = {
 	.resource	= pmu_resources,
 };
 
+/* an IPMMU module for ICB */
+static struct resource ipmmu_resources[] = {
+	[0] = {
+		.name	= "IPMMU",
+		.start	= 0xfe951000,
+		.end	= 0xfe9510ff,
+		.flags	= IORESOURCE_MEM,
+	},
+};
+
+static const char * const ipmmu_dev_names[] = {
+	"sh_mobile_lcdc_fb.0",
+};
+
+static struct shmobile_ipmmu_platform_data ipmmu_platform_data = {
+	.dev_names = ipmmu_dev_names,
+	.num_dev_names = ARRAY_SIZE(ipmmu_dev_names),
+};
+
+static struct platform_device ipmmu_device = {
+	.name           = "ipmmu",
+	.id             = -1,
+	.dev = {
+		.platform_data = &ipmmu_platform_data,
+	},
+	.resource       = ipmmu_resources,
+	.num_resources  = ARRAY_SIZE(ipmmu_resources),
+};
+
 static struct platform_device *sh73a0_early_devices[] __initdata = {
 	&scif0_device,
 	&scif1_device,
@@ -767,6 +797,7 @@ static struct platform_device *sh73a0_early_devices[] __initdata = {
 	&cmt10_device,
 	&tmu00_device,
 	&tmu01_device,
+	&ipmmu_device,
 };
 
 static struct platform_device *sh73a0_late_devices[] __initdata = {
-- 
1.8.0

[PATCH v5 4/4] ARM: mach-shmobile: r8a7740: Add IPMMU device

[Hideki EIRAKU]

This patch adds an IPMMU device and notifies the IPMMU driver which
devices are connected via the IPMMU module.  All devices connected to the main
memory bus via the IPMMU module MUST be registered when SHMOBILE_IPMMU and
SHMOBILE_IOMMU are enabled because physical address cannot be used
while the IPMMU module's MMU function is enabled.

Signed-off-by: Hideki EIRAKU <hdk@igel.co.jp>
---
 arch/arm/mach-shmobile/setup-r8a7740.c | 33 +++++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/arch/arm/mach-shmobile/setup-r8a7740.c b/arch/arm/mach-shmobile/setup-r8a7740.c
index 0952224..b85bea5 100644
--- a/arch/arm/mach-shmobile/setup-r8a7740.c
+++ b/arch/arm/mach-shmobile/setup-r8a7740.c
@@ -28,6 +28,7 @@
 #include <linux/sh_dma.h>
 #include <linux/sh_timer.h>
 #include <linux/dma-mapping.h>
+#include <linux/platform_data/sh_ipmmu.h>
 #include <mach/dma-register.h>
 #include <mach/r8a7740.h>
 #include <mach/pm-rmobile.h>
@@ -262,6 +263,37 @@ static struct platform_device cmt10_device = {
 	.num_resources	= ARRAY_SIZE(cmt10_resources),
 };
 
+/* IPMMUI (an IPMMU module for ICB/LMB) */
+static struct resource ipmmu_resources[] = {
+	[0] = {
+		.name	= "IPMMUI",
+		.start	= 0xfe951000,
+		.end	= 0xfe9510ff,
+		.flags	= IORESOURCE_MEM,
+	},
+};
+
+static const char * const ipmmu_dev_names[] = {
+	"sh_mobile_lcdc_fb.0",
+	"sh_mobile_lcdc_fb.1",
+	"sh_mobile_ceu.0",
+};
+
+static struct shmobile_ipmmu_platform_data ipmmu_platform_data = {
+	.dev_names = ipmmu_dev_names,
+	.num_dev_names = ARRAY_SIZE(ipmmu_dev_names),
+};
+
+static struct platform_device ipmmu_device = {
+	.name           = "ipmmu",
+	.id             = -1,
+	.dev = {
+		.platform_data = &ipmmu_platform_data,
+	},
+	.resource       = ipmmu_resources,
+	.num_resources  = ARRAY_SIZE(ipmmu_resources),
+};
+
 static struct platform_device *r8a7740_early_devices[] __initdata = {
 	&scif0_device,
 	&scif1_device,
@@ -273,6 +305,7 @@ static struct platform_device *r8a7740_early_devices[] __initdata = {
 	&scif7_device,
 	&scifb_device,
 	&cmt10_device,
+	&ipmmu_device,
 };
 
 /* DMA */
-- 
1.8.0

[PATCH v5 1/4] iommu/shmobile: Add iommu driver for Renesas IPMMU modules

[Laurent Pinchart]

Hi Eiraku-san,

Thank you for the patch. This version is definitely better than the previous 
one, but I still have a couple of comments.

On Tuesday 25 December 2012 20:19:12 Hideki EIRAKU wrote:
> This is the Renesas IPMMU driver and IOMMU API implementation.
> 
> The IPMMU module supports the MMU function and the PMB function.  The
> MMU function provides address translation by pagetable compatible with
> ARMv6.  The PMB function provides address translation including
> tile-linear translation.  This patch implements the MMU function.
> 
> The iommu driver does not register a platform driver directly because:
> - the register space of the MMU function and the PMB function
>   have a common register (used for settings flush), so they should ideally
>   have a way to appropriately share this register.
> - the MMU function uses the IOMMU API while the PMB function does not.
> - the two functions may be used independently.
> 
> Signed-off-by: Hideki EIRAKU <hdk@igel.co.jp>
> ---
>  drivers/iommu/Kconfig                  |  74 +++++++
>  drivers/iommu/Makefile                 |   2 +
>  drivers/iommu/shmobile-iommu.c         | 341 ++++++++++++++++++++++++++++++
>  drivers/iommu/shmobile-ipmmu.c         | 136 +++++++++++++
>  drivers/iommu/shmobile-ipmmu.h         |  37 ++++
>  include/linux/platform_data/sh_ipmmu.h |  18 ++
>  6 files changed, 608 insertions(+)
>  create mode 100644 drivers/iommu/shmobile-iommu.c
>  create mode 100644 drivers/iommu/shmobile-ipmmu.c
>  create mode 100644 drivers/iommu/shmobile-ipmmu.h
>  create mode 100644 include/linux/platform_data/sh_ipmmu.h
> 
> diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
> index e39f9db..d364494 100644
> --- a/drivers/iommu/Kconfig
> +++ b/drivers/iommu/Kconfig
> @@ -187,4 +187,78 @@ config EXYNOS_IOMMU_DEBUG
> 
>  	  Say N unless you need kernel log message for IOMMU debugging
> 
> +config SHMOBILE_IPMMU
> +	bool
> +
> +config SHMOBILE_IPMMU_TLB
> +	bool
> +
> +config SHMOBILE_IOMMU
> +	bool "IOMMU for Renesas IPMMU/IPMMUI"
> +	default n
> +	depends on (ARM && ARCH_SHMOBILE)
> +	select IOMMU_API
> +	select ARM_DMA_USE_IOMMU
> +	select SHMOBILE_IPMMU
> +	select SHMOBILE_IPMMU_TLB
> +	help
> +	  Support for Renesas IPMMU/IPMMUI. This option enables
> +	  remapping of DMA memory accesses from all of the IP blocks
> +	  on the ICB.
> +
> +	  Warning: Drivers (including userspace drivers of UIO
> +	  devices) of the IP blocks on the ICB *must* use addresses
> +	  allocated from the IPMMU (iova) for DMA with this option
> +	  enabled.
> +
> +	  If unsure, say N.
> +
> +choice
> +	prompt "IPMMU/IPMMUI address space size"
> +	default SHMOBILE_IOMMU_ADDRSIZE_2048MB
> +	depends on SHMOBILE_IOMMU
> +	help
> +	  This option sets IPMMU/IPMMUI address space size by
> +	  adjusting the 1st level page table size. The page table size
> +	  is calculated as follows:
> +
> +	      page table size = number of page table entries * 4 bytes
> +	      number of page table entries = address space size / 1 MiB
> +
> +	  For example, when the address space size is 2048 MiB, the
> +	  1st level page table size is 8192 bytes.
> +
> +	config SHMOBILE_IOMMU_ADDRSIZE_2048MB
> +		bool "2 GiB"
> +
> +	config SHMOBILE_IOMMU_ADDRSIZE_1024MB
> +		bool "1 GiB"
> +
> +	config SHMOBILE_IOMMU_ADDRSIZE_512MB
> +		bool "512 MiB"
> +
> +	config SHMOBILE_IOMMU_ADDRSIZE_256MB
> +		bool "256 MiB"
> +
> +	config SHMOBILE_IOMMU_ADDRSIZE_128MB
> +		bool "128 MiB"
> +
> +	config SHMOBILE_IOMMU_ADDRSIZE_64MB
> +		bool "64 MiB"
> +
> +	config SHMOBILE_IOMMU_ADDRSIZE_32MB
> +		bool "32 MiB"
> +
> +endchoice
> +
> +config SHMOBILE_IOMMU_L1SIZE
> +	int
> +	default 8192 if SHMOBILE_IOMMU_ADDRSIZE_2048MB
> +	default 4096 if SHMOBILE_IOMMU_ADDRSIZE_1024MB
> +	default 2048 if SHMOBILE_IOMMU_ADDRSIZE_512MB
> +	default 1024 if SHMOBILE_IOMMU_ADDRSIZE_256MB
> +	default 512 if SHMOBILE_IOMMU_ADDRSIZE_128MB
> +	default 256 if SHMOBILE_IOMMU_ADDRSIZE_64MB
> +	default 128 if SHMOBILE_IOMMU_ADDRSIZE_32MB
> +
>  endif # IOMMU_SUPPORT
> diff --git a/drivers/iommu/Makefile b/drivers/iommu/Makefile
> index f66b816..ef0e520 100644
> --- a/drivers/iommu/Makefile
> +++ b/drivers/iommu/Makefile
> @@ -13,3 +13,5 @@ obj-$(CONFIG_OMAP_IOMMU_DEBUG) += omap-iommu-debug.o
>  obj-$(CONFIG_TEGRA_IOMMU_GART) += tegra-gart.o
>  obj-$(CONFIG_TEGRA_IOMMU_SMMU) += tegra-smmu.o
>  obj-$(CONFIG_EXYNOS_IOMMU) += exynos-iommu.o
> +obj-$(CONFIG_SHMOBILE_IOMMU) += shmobile-iommu.o
> +obj-$(CONFIG_SHMOBILE_IPMMU) += shmobile-ipmmu.o
> diff --git a/drivers/iommu/shmobile-iommu.c b/drivers/iommu/shmobile-iommu.c
> new file mode 100644
> index 0000000..360a06f
> --- /dev/null
> +++ b/drivers/iommu/shmobile-iommu.c
> @@ -0,0 +1,341 @@
> +/*
> + * IOMMU for IPMMU/IPMMUI
> + * Copyright (C) 2012  Hideki EIRAKU
> + *
> + * 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; version 2 of the License.
> + */
> +
> +#include <linux/dma-mapping.h>
> +#include <linux/dmapool.h>
> +#include <linux/io.h>
> +#include <linux/iommu.h>
> +#include <linux/platform_device.h>
> +#include <linux/slab.h>
> +#include <asm/dma-iommu.h>
> +#include "shmobile-ipmmu.h"
> +
> +#define L1_SIZE CONFIG_SHMOBILE_IOMMU_L1SIZE
> +#define L1_LEN (L1_SIZE / 4)
> +#define L1_ALIGN L1_SIZE
> +#define L2_SIZE 0x400
> +#define L2_LEN (L2_SIZE / 4)
> +#define L2_ALIGN L2_SIZE
> +
> +struct shmobile_iommu_domain_pgtable {
> +	uint32_t *pgtable;
> +	dma_addr_t handle;
> +};
> +
> +struct shmobile_iommu_domain {
> +	struct shmobile_ipmmu *ipmmu;
> +	struct shmobile_iommu_domain_pgtable l1, l2[L1_LEN];
> +	spinlock_t map_lock;
> +	atomic_t active;
> +};
> +
> +static struct shmobile_ipmmu *platform_bus_ipmmu;
> +static struct dma_pool *l1pool, *l2pool;
> +static spinlock_t lock;
> +static struct shmobile_iommu_domain *attached;
> +static int num_attached_devices;

We need to get rid of those global variables. Most of them could probably be 
stored in the shmobile_ipmmu structure (or another IOMMU-specific structure 
emebdded in shmobile_ipmmu, or referenced from shmobile_ipmmu). Otherwise it 
won't be possible to instantiate more than on IPMMU in the system.

platform_bus_ipmmu will need special care, the driver will need to be 
restructured in such a way that no global IPMMU pointer is needed. Feel free 
to discuss your ideas on the mailing list before implementing them (and please 
CC me in that case).

> +static int shmobile_iommu_domain_init(struct iommu_domain *domain)
> +{
> +	struct shmobile_iommu_domain *sh_domain;
> +	int i;
> +
> +	sh_domain = kmalloc(sizeof(*sh_domain), GFP_KERNEL);
> +	if (!sh_domain)
> +		return -ENOMEM;
> +	sh_domain->l1.pgtable = dma_pool_alloc(l1pool, GFP_KERNEL,
> +					       &sh_domain->l1.handle);
> +	if (!sh_domain->l1.pgtable) {
> +		kfree(sh_domain);
> +		return -ENOMEM;
> +	}
> +	for (i = 0; i < L1_LEN; i++)
> +		sh_domain->l2[i].pgtable = NULL;
> +	memset(sh_domain->l1.pgtable, 0, L1_SIZE);
> +	spin_lock_init(&sh_domain->map_lock);
> +	atomic_set(&sh_domain->active, 0);
> +	domain->priv = sh_domain;
> +	return 0;
> +}
> +
> +static void shmobile_iommu_domain_destroy(struct iommu_domain *domain)
> +{
> +	struct shmobile_iommu_domain *sh_domain = domain->priv;
> +	int i;
> +
> +	for (i = 0; i < L1_LEN; i++) {
> +		if (sh_domain->l2[i].pgtable)
> +			dma_pool_free(l2pool, sh_domain->l2[i].pgtable,
> +				      sh_domain->l2[i].handle);
> +	}
> +	dma_pool_free(l1pool, sh_domain->l1.pgtable, sh_domain->l1.handle);
> +	kfree(sh_domain);
> +	domain->priv = NULL;
> +}
> +
> +static int shmobile_iommu_attach_device(struct iommu_domain *domain,
> +					struct device *dev)
> +{
> +	struct shmobile_iommu_domain *sh_domain = domain->priv;
> +	struct shmobile_ipmmu *ipmmu = platform_bus_ipmmu;
> +	int ret = -EBUSY;
> +
> +	spin_lock(&lock);
> +	if (attached != sh_domain) {
> +		if (attached)
> +			goto err;
> +		atomic_set(&sh_domain->active, 1);
> +		ipmmu_tlb_set(ipmmu, sh_domain->l1.handle, L1_SIZE, 0);
> +		wmb();
> +		ipmmu_tlb_flush(ipmmu);
> +		attached = sh_domain;
> +		num_attached_devices = 0;
> +		sh_domain->ipmmu = ipmmu;
> +	}
> +	num_attached_devices++;
> +	ret = 0;
> +err:
> +	spin_unlock(&lock);
> +	return ret;
> +}
> +
> +static void shmobile_iommu_detach_device(struct iommu_domain *domain,
> +					 struct device *dev)
> +{
> +	struct shmobile_iommu_domain *sh_domain = domain->priv;
> +	struct shmobile_ipmmu *ipmmu = platform_bus_ipmmu;
> +
> +	spin_lock(&lock);
> +	atomic_set(&sh_domain->active, 0);
> +	num_attached_devices--;
> +	if (!num_attached_devices) {
> +		ipmmu_tlb_set(ipmmu, 0, 0, 0);
> +		ipmmu_tlb_flush(ipmmu);
> +		attached = NULL;
> +		sh_domain->ipmmu = NULL;
> +	}
> +	spin_unlock(&lock);
> +}
> +
> +static int
> +l2alloc(struct shmobile_iommu_domain *sh_domain, unsigned int l1index)
> +{
> +	if (!sh_domain->l2[l1index].pgtable) {
> +		sh_domain->l2[l1index].pgtable = dma_pool_alloc(l2pool,
> +						GFP_KERNEL,
> +						&sh_domain->l2[l1index].handle);
> +		if (!sh_domain->l2[l1index].pgtable)
> +			return -ENOMEM;
> +		memset(sh_domain->l2[l1index].pgtable, 0, L2_SIZE);
> +	}
> +	sh_domain->l1.pgtable[l1index] = sh_domain->l2[l1index].handle | 0x1;
> +	return 0;
> +}
> +
> +static void
> +l2realfree(struct shmobile_iommu_domain_pgtable *l2)
> +{
> +	if (l2->pgtable)
> +		dma_pool_free(l2pool, l2->pgtable, l2->handle);
> +}
> +
> +static int
> +l2free(struct shmobile_iommu_domain *sh_domain, unsigned int l1index,
> +	struct shmobile_iommu_domain_pgtable *l2)
> +{
> +	sh_domain->l1.pgtable[l1index] = 0;
> +	if (sh_domain->l2[l1index].pgtable) {
> +		*l2 = sh_domain->l2[l1index];
> +		sh_domain->l2[l1index].pgtable = NULL;
> +	}
> +	return 0;
> +}
> +
> +static int shmobile_iommu_map(struct iommu_domain *domain, unsigned long
> iova,
> +			      phys_addr_t paddr, size_t size, int prot)
> +{
> +	struct shmobile_iommu_domain_pgtable l2 = { .pgtable = NULL };
> +	struct shmobile_iommu_domain *sh_domain = domain->priv;
> +	unsigned int l1index, l2index, i;
> +	int ret;
> +
> +	l1index = iova >> 20;
> +	switch (size) {
> +	case 0x1000:

Could you please use the macros defines in include/linux/sizes.h ? They make 
the code easier to read.

> +		l2index = (iova >> 12) & 0xff;
> +		spin_lock(&sh_domain->map_lock);
> +		ret = l2alloc(sh_domain, l1index);

l2alloc calls dma_pool_alloc(GFP_KERNEL), that not safe in a non-sleepable 
context. Do we need a spinlock here, or could a mutex do ?

> +		if (!ret)
> +			sh_domain->l2[l1index].pgtable[l2index] = paddr | 0xff2;
> +		spin_unlock(&sh_domain->map_lock);
> +		break;
> +	case 0x10000:
> +		l2index = (iova >> 12) & 0xf0;
> +		spin_lock(&sh_domain->map_lock);
> +		ret = l2alloc(sh_domain, l1index);
> +		if (!ret) {
> +			for (i = 0; i < 0x10; i++)
> +				sh_domain->l2[l1index].pgtable[l2index + i] =
> +					paddr | 0xff1;
> +		}
> +		spin_unlock(&sh_domain->map_lock);
> +		break;
> +	case 0x100000:
> +		spin_lock(&sh_domain->map_lock);
> +		l2free(sh_domain, l1index, &l2);
> +		sh_domain->l1.pgtable[l1index] = paddr | 0xc02;
> +		spin_unlock(&sh_domain->map_lock);
> +		ret = 0;
> +		break;
> +	default:
> +		ret = -EINVAL;
> +	}
> +	if (!ret && atomic_read(&sh_domain->active)) {
> +		wmb();
> +		ipmmu_tlb_flush(sh_domain->ipmmu);
> +		l2realfree(&l2);

Does l2realfree() need locking ?

> +	}
> +	return ret;
> +}
> +
> +static size_t shmobile_iommu_unmap(struct iommu_domain *domain,
> +				   unsigned long iova, size_t size)
> +{
> +	struct shmobile_iommu_domain_pgtable l2 = { .pgtable = NULL };
> +	struct shmobile_iommu_domain *sh_domain = domain->priv;
> +	unsigned int l1index, l2index, i;
> +	uint32_t l2entry = 0;
> +	size_t ret = 0;
> +
> +	l1index = iova >> 20;
> +	if (!(iova & 0xFFFFF) && size >= 0x100000) {

Hex constants usually use lower-case in the kernel.

Could you please use SZ_* for the size here (and below, as applicable) as well 
?

> +		spin_lock(&sh_domain->map_lock);
> +		l2free(sh_domain, l1index, &l2);
> +		spin_unlock(&sh_domain->map_lock);
> +		ret = 0x100000;
> +		goto done;
> +	}
> +	l2index = (iova >> 12) & 0xff;
> +	spin_lock(&sh_domain->map_lock);
> +	if (sh_domain->l2[l1index].pgtable)
> +		l2entry = sh_domain->l2[l1index].pgtable[l2index];
> +	switch (l2entry & 3) {
> +	case 1:
> +		if (l2index & 0xf)
> +			break;
> +		for (i = 0; i < 0x10; i++)
> +			sh_domain->l2[l1index].pgtable[l2index + i] = 0;
> +		ret = 0x10000;
> +		break;
> +	case 2:
> +		sh_domain->l2[l1index].pgtable[l2index] = 0;
> +		ret = 0x1000;
> +		break;
> +	}
> +	spin_unlock(&sh_domain->map_lock);
> +done:
> +	if (ret && atomic_read(&sh_domain->active)) {
> +		wmb();
> +		ipmmu_tlb_flush(sh_domain->ipmmu);
> +		l2realfree(&l2);
> +	}
> +	return ret;
> +}
> +
> +static phys_addr_t shmobile_iommu_iova_to_phys(struct iommu_domain *domain,
> +					       unsigned long iova)
> +{
> +	struct shmobile_iommu_domain *sh_domain = domain->priv;
> +	uint32_t l1entry = 0, l2entry = 0;
> +	unsigned int l1index, l2index;
> +
> +	l1index = iova >> 20;
> +	l2index = (iova >> 12) & 0xff;
> +	spin_lock(&sh_domain->map_lock);
> +	if (sh_domain->l2[l1index].pgtable)
> +		l2entry = sh_domain->l2[l1index].pgtable[l2index];
> +	else
> +		l1entry = sh_domain->l1.pgtable[l1index];
> +	spin_unlock(&sh_domain->map_lock);
> +	switch (l2entry & 3) {
> +	case 1:
> +		return (l2entry & ~0xffff) | (iova & 0xffff);
> +	case 2:
> +		return (l2entry & ~0xfff) | (iova & 0xfff);
> +	default:
> +		if ((l1entry & 3) == 2)
> +			return (l1entry & ~0xfffff) | (iova & 0xfffff);
> +		return 0;
> +	}
> +}
> +
> +static int find_dev_name(struct shmobile_ipmmu *ipmmu, const char
> *dev_name) +{
> +	unsigned int i, n = ipmmu->num_dev_names;
> +
> +	for (i = 0; i < n; i++) {
> +		if (strcmp(ipmmu->dev_names[i], dev_name) == 0)
> +			return 1;
> +	}
> +	return 0;
> +}
> +
> +static int shmobile_iommu_add_device(struct device *dev)
> +{
> +	struct shmobile_ipmmu *ipmmu = platform_bus_ipmmu;
> +	struct dma_iommu_mapping *mapping;
> +
> +	if (!find_dev_name(ipmmu, dev_name(dev)))
> +		return 0;
> +	mapping = ipmmu->iommu_mapping;
> +	if (!mapping) {
> +		mapping = arm_iommu_create_mapping(&platform_bus_type, 0,
> +						   L1_LEN << 20, 0);
> +		if (IS_ERR(mapping))
> +			return PTR_ERR(mapping);
> +		ipmmu->iommu_mapping = mapping;
> +	}
> +	if (arm_iommu_attach_device(dev, mapping))
> +		pr_err("arm_iommu_attach_device failed\n");

This assumes that the IPMMU is registered before the device using it. I don't 
think that will hold true with DT.

As stated in my previous e-mail, I really think you should continue the 
discussion started by patch http://patchwork.ozlabs.org/patch/203717/ on the 
mailing lists (please CC me).

> +	return 0;
> +}
> +
> +static struct iommu_ops shmobile_iommu_ops = {
> +	.domain_init = shmobile_iommu_domain_init,
> +	.domain_destroy = shmobile_iommu_domain_destroy,
> +	.attach_dev = shmobile_iommu_attach_device,
> +	.detach_dev = shmobile_iommu_detach_device,
> +	.map = shmobile_iommu_map,
> +	.unmap = shmobile_iommu_unmap,
> +	.iova_to_phys = shmobile_iommu_iova_to_phys,
> +	.add_device = shmobile_iommu_add_device,
> +	.pgsize_bitmap = 0x111000,
> +};
> +
> +int ipmmu_iommu_init(struct shmobile_ipmmu *ipmmu)
> +{
> +	dma_set_coherent_mask(ipmmu->dev, DMA_BIT_MASK(32));
> +	l1pool = dma_pool_create("shmobile-iommu-pgtable1", ipmmu->dev,
> +				 L1_SIZE, L1_ALIGN, 0);
> +	if (!l1pool)
> +		goto nomem_pool1;

You can return -ENOMEM directly here.

> +	l2pool = dma_pool_create("shmobile-iommu-pgtable2", ipmmu->dev,
> +				 L2_SIZE, L2_ALIGN, 0);
> +	if (!l2pool)
> +		goto nomem_pool2;

And you can call dma_pool_destroy() and return -ENOMEM directly here too, as 
there no other user of the error code path.

> +	spin_lock_init(&lock);
> +	attached = NULL;
> +	platform_bus_ipmmu = ipmmu;
> +	bus_set_iommu(&platform_bus_type, &shmobile_iommu_ops);
> +	return 0;
> +nomem_pool2:
> +	dma_pool_destroy(l1pool);
> +nomem_pool1:
> +	return -ENOMEM;
> +}
> diff --git a/drivers/iommu/shmobile-ipmmu.c b/drivers/iommu/shmobile-ipmmu.c
> new file mode 100644
> index 0000000..8321f89
> --- /dev/null
> +++ b/drivers/iommu/shmobile-ipmmu.c
> @@ -0,0 +1,136 @@
> +/*
> + * IPMMU/IPMMUI
> + * Copyright (C) 2012  Hideki EIRAKU
> + *
> + * 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; version 2 of the License.
> + */
> +
> +#include <linux/err.h>
> +#include <linux/export.h>
> +#include <linux/io.h>
> +#include <linux/platform_device.h>
> +#include <linux/slab.h>
> +#include <linux/platform_data/sh_ipmmu.h>
> +#include "shmobile-ipmmu.h"
> +
> +#define IMCTR1 0x000
> +#define IMCTR2 0x004
> +#define IMASID 0x010
> +#define IMTTBR 0x014
> +#define IMTTBCR 0x018
> +
> +#define IMCTR1_TLBEN (1 << 0)
> +#define IMCTR1_FLUSH (1 << 1)
> +
> +static void ipmmu_reg_write(struct shmobile_ipmmu *ipmmu, unsigned long
> reg_off,
> +			    unsigned long data)
> +{
> +	iowrite32(data, ipmmu->ipmmu_base + reg_off);
> +}
> +
> +void ipmmu_tlb_flush(struct shmobile_ipmmu *ipmmu)
> +{
> +	if (!ipmmu)
> +		return;
> +
> +	mutex_lock(&ipmmu->flush_lock);
> +	if (ipmmu->tlb_enabled)
> +		ipmmu_reg_write(ipmmu, IMCTR1, IMCTR1_FLUSH | IMCTR1_TLBEN);
> +	else
> +		ipmmu_reg_write(ipmmu, IMCTR1, IMCTR1_FLUSH);
> +	mutex_unlock(&ipmmu->flush_lock);
> +}
> +
> +void ipmmu_tlb_set(struct shmobile_ipmmu *ipmmu, unsigned long phys, int
> size,
> +		   int asid)
> +{
> +	if (!ipmmu)
> +		return;
> +
> +	mutex_lock(&ipmmu->flush_lock);
> +	switch (size) {
> +	default:
> +		ipmmu->tlb_enabled = 0;
> +		break;
> +	case 0x2000:
> +		ipmmu_reg_write(ipmmu, IMTTBCR, 1);
> +		ipmmu->tlb_enabled = 1;
> +		break;
> +	case 0x1000:
> +		ipmmu_reg_write(ipmmu, IMTTBCR, 2);
> +		ipmmu->tlb_enabled = 1;
> +		break;
> +	case 0x800:
> +		ipmmu_reg_write(ipmmu, IMTTBCR, 3);
> +		ipmmu->tlb_enabled = 1;
> +		break;
> +	case 0x400:
> +		ipmmu_reg_write(ipmmu, IMTTBCR, 4);
> +		ipmmu->tlb_enabled = 1;
> +		break;
> +	case 0x200:
> +		ipmmu_reg_write(ipmmu, IMTTBCR, 5);
> +		ipmmu->tlb_enabled = 1;
> +		break;
> +	case 0x100:
> +		ipmmu_reg_write(ipmmu, IMTTBCR, 6);
> +		ipmmu->tlb_enabled = 1;
> +		break;
> +	case 0x80:
> +		ipmmu_reg_write(ipmmu, IMTTBCR, 7);
> +		ipmmu->tlb_enabled = 1;
> +		break;
> +	}
> +	ipmmu_reg_write(ipmmu, IMTTBR, phys);
> +	ipmmu_reg_write(ipmmu, IMASID, asid);
> +	mutex_unlock(&ipmmu->flush_lock);
> +}
> +
> +static int ipmmu_probe(struct platform_device *pdev)
> +{
> +	struct shmobile_ipmmu *ipmmu;
> +	struct resource *res;
> +	struct shmobile_ipmmu_platform_data *pdata = pdev->dev.platform_data;
> +
> +	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> +	if (!res) {
> +		dev_err(&pdev->dev, "cannot get platform resources\n");
> +		return -ENOENT;
> +	}
> +	ipmmu = devm_kzalloc(&pdev->dev, sizeof(*ipmmu), GFP_KERNEL);
> +	if (!ipmmu) {
> +		dev_err(&pdev->dev, "cannot allocate device data\n");
> +		return -ENOMEM;
> +	}
> +	mutex_init(&ipmmu->flush_lock);
> +	ipmmu->dev = &pdev->dev;
> +	ipmmu->ipmmu_base = devm_ioremap_nocache(&pdev->dev, res->start,
> +						resource_size(res));
> +	if (!ipmmu->ipmmu_base) {
> +		dev_err(&pdev->dev, "ioremap_nocache failed\n");
> +		return -ENOMEM;
> +	}
> +	ipmmu->dev_names = pdata->dev_names;
> +	ipmmu->num_dev_names = pdata->num_dev_names;
> +	platform_set_drvdata(pdev, ipmmu);
> +	ipmmu_reg_write(ipmmu, IMCTR1, 0x0); /* disable TLB */
> +	ipmmu_reg_write(ipmmu, IMCTR2, 0x0); /* disable PMB */
> +	ipmmu_iommu_init(ipmmu);
> +	return 0;
> +}
> +
> +static struct platform_driver ipmmu_driver = {
> +	.probe = ipmmu_probe,
> +	.driver = {
> +		.owner = THIS_MODULE,
> +		.name = "ipmmu",
> +	},
> +};
> +
> +static int __init ipmmu_init(void)
> +{
> +	return platform_driver_register(&ipmmu_driver);
> +}
> +subsys_initcall(ipmmu_init);
> diff --git a/drivers/iommu/shmobile-ipmmu.h b/drivers/iommu/shmobile-ipmmu.h
> new file mode 100644
> index 0000000..6270e7c
> --- /dev/null
> +++ b/drivers/iommu/shmobile-ipmmu.h
> @@ -0,0 +1,37 @@
> +/* shmobile-ipmmu.h
> + *
> + * Copyright (C) 2012  Hideki EIRAKU
> + *
> + * 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; version 2 of the License.
> + */
> +
> +#ifndef __SHMOBILE_IPMMU_H__
> +#define __SHMOBILE_IPMMU_H__
> +
> +struct dma_iommu_mapping;
> +
> +struct shmobile_ipmmu {
> +	struct device *dev;
> +	void __iomem *ipmmu_base;
> +	int tlb_enabled;
> +	struct mutex flush_lock;
> +	struct dma_iommu_mapping *iommu_mapping;
> +	const char * const *dev_names;
> +	unsigned int num_dev_names;
> +};
> +
> +#ifdef CONFIG_SHMOBILE_IPMMU_TLB
> +void ipmmu_tlb_flush(struct shmobile_ipmmu *ipmmu);
> +void ipmmu_tlb_set(struct shmobile_ipmmu *ipmmu, unsigned long phys, int
> size, +		   int asid);
> +int ipmmu_iommu_init(struct shmobile_ipmmu *ipmmu);
> +#else
> +static int ipmmu_iommu_init(struct shmobile_ipmmu *ipmmu)
> +{
> +	return -EINVAL;
> +}
> +#endif
> +
> +#endif /* __SHMOBILE_IPMMU_H__ */
> diff --git a/include/linux/platform_data/sh_ipmmu.h
> b/include/linux/platform_data/sh_ipmmu.h new file mode 100644
> index 0000000..39f7405
> --- /dev/null
> +++ b/include/linux/platform_data/sh_ipmmu.h
> @@ -0,0 +1,18 @@
> +/* sh_ipmmu.h
> + *
> + * Copyright (C) 2012  Hideki EIRAKU
> + *
> + * 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; version 2 of the License.
> + */
> +
> +#ifndef __SH_IPMMU_H__
> +#define __SH_IPMMU_H__
> +
> +struct shmobile_ipmmu_platform_data {
> +	const char * const *dev_names;
> +	unsigned int num_dev_names;
> +};
> +
> +#endif /* __SH_IPMMU_H__ */
-- 
Regards,

Laurent Pinchart

[PATCH v5 1/4] iommu/shmobile: Add iommu driver for Renesas IPMMU modules

[Joerg Roedel]

Hi,

On Mon, Jan 07, 2013 at 07:11:58PM +0100, Laurent Pinchart wrote:
> > +		l2index = (iova >> 12) & 0xff;
> > +		spin_lock(&sh_domain->map_lock);
> > +		ret = l2alloc(sh_domain, l1index);
> 
> l2alloc calls dma_pool_alloc(GFP_KERNEL), that not safe in a non-sleepable 
> context. Do we need a spinlock here, or could a mutex do ?

iommu_map should work in any context, so a mutex will not work. Also the
memory allocations in that path should be GFP_ATOMIC instead of
GFP_KERNEL.

Other than that this driver looks good from an IOMMU-API perspective.
Please Cc me on future versions of this patch-set directly.

Thanks,

	Joerg