[ARM] tegra: add I/O virtual memory manager interface (iovmm)
authorGary King <gking@nvidia.com>
Sat, 26 Jun 2010 01:39:58 +0000 (18:39 -0700)
committerColin Cross <ccross@android.com>
Wed, 6 Oct 2010 23:27:10 +0000 (16:27 -0700)
The Tegra IOVMM is an interface to allow device drivers and subsystems in
the kernel to manage the virtual memory spaces visible to I/O devices.

The interface has been designed to be scalable to allow for I/O virtual
memory hardware which exists in one or more limited apertures of the address
space (e.g., a small aperture in physical address space which can perform
MMU-like remapping) up to complete virtual addressing with multiple
address spaces and memory protection.

The interface has been designed to be similar to the Linux virtual memory
system; however, operations which would be difficult to implement or
nonsensical for DMA devices (e.g., copy-on-write) are not present, and
APIs have been added to allow for management of multiple simultaneous
active address spaces.

The API is broken into four principal objects: areas, clients, domains and
devices.

Areas
=====

An area is a contiguous region of the virtual address space which can be
filled with virtual-to-physical translations (and, optionally, protection
attributes). The virtual address of the area can be queried and used for
DMA operations by the client which created it.

As with the Linux vm_area structures, it is the responsibility of whichever
code creates an area to ensure that it is populated with appropriate
translations.

Domains
=======

A domain in the IOVMM system is similar to a process in a standard CPU
virtual memory system; it represents the entire range of virtual addresses
which may be allocated and used for translation. Depending on hardware
capabilities, one or more domains may be resident and available for
translation. IOVMM areas are allocated from IOVMM domains.

Whenever a DMA operation is performed to or from an IOVMM area, its parent
domain must be made resident prior to commencing the operation.

Clients
=======

I/O VMM clients represent any entity which needs to be able to allocate
and map system memory into I/O virtual space. Clients are created by name
and may be created as part of a "share group," where all clients created
in the same share group will observe the same I/O virtual space (i.e., all
will use the same IOVMM domain). This is similar to threads inside a process
in the CPU virtual memory manager.

The callers of the I/O VMM system are responsible for deciding on the
granularity of client creation and share group definition; depending on the
specific usage model expected by the caller, it may be appropriate to create
an IOVMM client per task (if the caller represents an ioctl'able interface
to user land), an IOVMM client per driver instance, a common IOVMM client
for an entire bus, or a global IOVMM client for an OS subsystem (e.g., the DMA
mapping interface).

Each client is responsible for ensuring that its IOVMM client's translation is
resident on the system prior to performing DMA operations using the IOVMM
addresses. This is accomplished by preceding all DMA operations for the client
with a call to tegra_iovmm_client_lock (or tegra_iovmm_client_trylock),
and following all operations (once complete) with a call to
tegra_iovmm_client_unlock. In this regard, clients are cooperatively context-
switched, and are expected to behave appropriately.

Devices
=======

I/O VMM devices are the physical hardware which is responsible for performing
the I/O virtual-to-physical translation.

Devices are responsible for domain management: the mapping and unmapping
operations needed to make translations resident in the domain (including
any TLB shootdown or cache invalidation needed to ensure coherency), locking
and unlocking domains as they are made resident by clients into the devices'
address space(s), and allocating and deallocating the domain objects.

Devices are responsible for the allocation and deallocation of domains to
allow coalescing of multiple client share groups into a single domain. For
example, if the device's hardware only allows a single address space to
be translated system-wide, performing full flushes and invalidates of the
translation at every client switch may be prohibitively expensive. In these
circumstances, a legal implementation of the IOVMM interface includes
returning the same domain for all clients on the system (regardless of
the originally-specified share group).

In this respect, a client can be assured that it will share an address space
with all of the other clients in its share group; however, it may also share
this address space with other clients, too.

Multiple devices may be present in a system; a device should return a NULL
domain if it is incapable of servicing the client when it is asked to
allocate a domain.

----------------------------------------------------------------------------

IOVMM Client API
================

tegra_iovmm_alloc_client - Called to create a new IOVMM client object; the
 implementation may create a new domain or return an existing one depending on
 both the device and the share group.

tegra_iovmm_free_client - Frees a client.

tegra_iovmm_client_lock - Makes a client's translations resident in the IOVMM
 device for subsequent DMA operations. May block if the device is incapable
 of context-switching the client when it is called. Returns -EINTR if the
 waiting thread is interrupted before the client is locked.

tegra_iovmm_client_trylock - Non-blocking version of tegra_iovmm_client_lock

tegra_iovmm_client_unlock - Called by clients after DMA operations on IOVMM-
 translated addresses is complete; allows IOVMM system to context-switch the
 current client out of the device if needed.

tegra_iovmm_create_vm - Called to allocate an IOVMM area. If
 lazy / demand-loading of pages is desired, clients should supply a pointer
 to a tegra_iovmm_area_ops structure providing callback functions to load, pin
 and unpin the physical pages which will be mapped into this IOVMM region.

tegra_iovmm_get_vm_size - Called to query the total size of an IOVMM client

tegra_iovmm_free_vm - Called to free a IOVMM area, releasing any pinned
 physical pages mapped by it and to decommit any resources (memory for
 PTEs / PDEs) required by the VM area.

tegra_iovmm_vm_insert_pfn - Called to insert an exact pfn (system memory
 physical page) into the area at a specific virtual address. Illegal to call
 if the IOVMM area was originally created with lazy / demand-loading.

tegra_iovmm_zap_vm - Called to mark all mappings in the IOVMM area as
 invalid / no-access, but continues to consume the I/O virtual address space.
 For lazy / demand-loaded IOVMM areas, a zapped region will not be reloaded
 until it has been unzapped; DMA operations using the affected translations
 may fault (if supported by the device).

tegra_iovmm_unzap_vm - Called to re-enable lazy / demand-loading of pages
 for a previously-zapped IOVMM area.

tegra_iovmm_find_area_get - Called to find the IOVMM area object
 corresponding to the specified I/O virtual address, or NULL if the address
 is not allocated in the client's address space. Increases the reference count
 on the IOVMM area object

tegra_iovmm_area_get - Called to increase the reference count on the IOVMM
 area object

tegra_iovmm_area_put - Called to decrease the reference count on the IOVMM
 area object

IOVMM Device API
================

tegra_iovmm_register - Called to register a new IOVMM device with the IOVMM
 manager

tegra_iovmm_unregister - Called to remove an IOVMM device from the IOVMM
 manager (unspecified behavior if called while a translation is active and / or
 in-use)

tegra_iovmm_domain_init - Called to initialize all of the IOVMM manager's
 data structures (block trees, etc.) after allocating a new domain

IOVMM Device HAL
================

map - Called to inform the device about a new lazy-mapped IOVMM area. Devices
 may load the entire VM area when this is called, or at any time prior to
 the completion of the first read or write operation using the translation.

unmap - Called to zap or to decommit translations

map_pfn - Called to insert a specific virtual-to-physical translation in the
 IOVMM area

lock_domain - Called to make a domain resident; should return 0 if the
 domain was successfully context-switched, non-zero if the operation can
 not be completed (e.g., all available simultaneous hardware translations are
 locked). If the device can guarantee that every domain it allocates is
 always usable, this function may be NULL.

unlock_domain - Releases a domain from residency, allows the hardware
 translation to be used by other domains.

alloc_domain - Called to allocate a new domain; allowed to return an
 existing domain

free_domain - Called to free a domain.

Change-Id: Ic65788777b7aba50ee323fe16fd553ce66c4b87c
Signed-off-by: Gary King <gking@nvidia.com>
arch/arm/mach-tegra/Kconfig
arch/arm/mach-tegra/Makefile
arch/arm/mach-tegra/include/mach/iovmm.h [new file with mode: 0644]
arch/arm/mach-tegra/iovmm-gart.c [new file with mode: 0644]
arch/arm/mach-tegra/iovmm.c [new file with mode: 0644]

index f33d011b16d383fb56d660e0a1780b5caf40a000..3eb494a7d1d5746f937d21469048948afd573b76 100644 (file)
@@ -56,3 +56,16 @@ config TEGRA_SYSTEM_DMA
          several Tegra device drivers
 
 endif
+
+config TEGRA_IOVMM_GART
+       bool "Enable I/O virtual memory manager for GART"
+       depends on ARCH_TEGRA_2x_SOC
+       default y
+       select TEGRA_IOVMM
+       help
+         Enables support for remapping discontiguous physical memory
+         shared with the operating system into contiguous I/O virtual
+         space through the GART hardware included on Tegra SoCs
+
+config TEGRA_IOVMM
+       bool
index 9c7cfd95373e6c0fc27eb99963c5eeed1eb62400..f4f18b7a45f5a3153706aebad57054fe4268dd36 100644 (file)
@@ -28,6 +28,8 @@ endif
 obj-$(CONFIG_TEGRA_SYSTEM_DMA)         += dma.o
 obj-$(CONFIG_CPU_FREQ)                  += cpu-tegra.o
 obj-$(CONFIG_CPU_IDLE)                 += cpuidle.o
+obj-$(CONFIG_TEGRA_IOVMM)               += iovmm.o
+obj-$(CONFIG_TEGRA_IOVMM_GART)          += iovmm-gart.o
 
 obj-${CONFIG_MACH_HARMONY}              += board-harmony.o
 obj-${CONFIG_MACH_HARMONY}              += board-harmony-pinmux.o
diff --git a/arch/arm/mach-tegra/include/mach/iovmm.h b/arch/arm/mach-tegra/include/mach/iovmm.h
new file mode 100644 (file)
index 0000000..8f11160
--- /dev/null
@@ -0,0 +1,286 @@
+/*
+ * arch/arm/mach-tegra/include/mach/iovmm.h
+ *
+ * Copyright (c) 2010, NVIDIA Corporation.
+ *
+ * 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 i 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.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
+ */
+
+#include <linux/list.h>
+#include <linux/platform_device.h>
+#include <linux/rbtree.h>
+#include <linux/rwsem.h>
+#include <linux/spinlock.h>
+#include <linux/types.h>
+
+#ifndef _MACH_TEGRA_IOVMM_H_
+#define _MACH_TEGRA_IOVMM_H_
+
+#if defined(CONFIG_ARCH_TEGRA_2x_SOC)
+typedef u32 tegra_iovmm_addr_t;
+#else
+#error "Unsupported tegra architecture family"
+#endif
+
+struct tegra_iovmm_device_ops;
+
+/* each I/O virtual memory manager unit should register a device with
+ * the iovmm system
+ */
+struct tegra_iovmm_device {
+       struct tegra_iovmm_device_ops   *ops;
+       const char                      *name;
+       struct list_head                list;
+       int                             pgsize_bits;
+};
+
+/* tegra_iovmm_domain serves a purpose analagous to mm_struct as defined in
+ * <linux/mm_types.h> - it defines a virtual address space within which
+ * tegra_iovmm_areas can be created.
+ */
+struct tegra_iovmm_domain {
+       atomic_t                clients;
+       atomic_t                locks;
+       spinlock_t              block_lock;
+       unsigned long           flags;
+       wait_queue_head_t       delay_lock;  /* when lock_client fails */
+       struct rw_semaphore     map_lock;
+       struct rb_root          all_blocks;  /* ordered by address */
+       struct rb_root          free_blocks; /* ordered by size */
+       struct tegra_iovmm_device *dev;
+};
+
+/* tegra_iovmm_client is analagous to an individual task in the task group
+ * which owns an mm_struct.
+ */
+
+struct iovmm_share_group;
+
+struct tegra_iovmm_client {
+       const char                      *name;
+       unsigned long                   flags;
+       struct iovmm_share_group        *group;
+       struct tegra_iovmm_domain       *domain;
+       struct list_head                list;
+};
+
+/* tegra_iovmm_area serves a purpose analagous to vm_area_struct as defined
+ * in <linux/mm_types.h> - it defines a virtual memory area which can be
+ * mapped to physical memory by a client-provided mapping function. */
+
+struct tegra_iovmm_area {
+       struct tegra_iovmm_domain       *domain;
+       tegra_iovmm_addr_t              iovm_start;
+       tegra_iovmm_addr_t              iovm_length;
+       pgprot_t                        pgprot;
+       struct tegra_iovmm_area_ops     *ops;
+};
+
+struct tegra_iovmm_device_ops {
+       /* maps a VMA using the page residency functions provided by the VMA */
+       int (*map)(struct tegra_iovmm_device *dev,
+               struct tegra_iovmm_area *io_vma);
+       /* marks all PTEs in a VMA as invalid; decommits the virtual addres
+        * space (potentially freeing PDEs when decommit is true.) */
+       void (*unmap)(struct tegra_iovmm_device *dev,
+               struct tegra_iovmm_area *io_vma, bool decommit);
+       void (*map_pfn)(struct tegra_iovmm_device *dev,
+               struct tegra_iovmm_area *io_vma,
+               tegra_iovmm_addr_t offs, unsigned long pfn);
+       /* ensures that a domain is resident in the hardware's mapping region
+        * so that it may be used by a client */
+       int (*lock_domain)(struct tegra_iovmm_device *dev,
+               struct tegra_iovmm_domain *domain);
+       void (*unlock_domain)(struct tegra_iovmm_device *dev,
+               struct tegra_iovmm_domain *domain);
+       /* allocates a vmm_domain for the specified client; may return the same
+        * domain for multiple clients */
+       struct tegra_iovmm_domain* (*alloc_domain)(
+               struct tegra_iovmm_device *dev,
+               struct tegra_iovmm_client *client);
+       void (*free_domain)(struct tegra_iovmm_device *dev,
+               struct tegra_iovmm_domain *domain);
+       int (*suspend)(struct tegra_iovmm_device *dev);
+       void (*resume)(struct tegra_iovmm_device *dev);
+};
+
+struct tegra_iovmm_area_ops {
+       /* ensures that the page of data starting at the specified offset
+        * from the start of the iovma is resident and pinned for use by
+        * DMA, returns the system pfn, or an invalid pfn if the
+        * operation fails. */
+       unsigned long (*lock_makeresident)(struct tegra_iovmm_area *area,
+               tegra_iovmm_addr_t offs);
+       /* called when the page is unmapped from the I/O VMA */
+       void (*release)(struct tegra_iovmm_area *area, tegra_iovmm_addr_t offs);
+};
+
+#ifdef CONFIG_TEGRA_IOVMM
+/* called by clients to allocate an I/O VMM client mapping context which
+ * will be shared by all clients in the same share_group */
+struct tegra_iovmm_client *tegra_iovmm_alloc_client(const char *name,
+       const char *share_group);
+
+size_t tegra_iovmm_get_vm_size(struct tegra_iovmm_client *client);
+
+void tegra_iovmm_free_client(struct tegra_iovmm_client *client);
+
+/* called by clients to ensure that their mapping context is resident
+ * before performing any DMA operations addressing I/O VMM regions.
+ * client_lock may return -EINTR. */
+int tegra_iovmm_client_lock(struct tegra_iovmm_client *client);
+int tegra_iovmm_client_trylock(struct tegra_iovmm_client *client);
+
+/* called by clients after DMA operations are complete */
+void tegra_iovmm_client_unlock(struct tegra_iovmm_client *client);
+
+/* called by clients to allocate a new iovmm_area and reserve I/O virtual
+ * address space for it. if ops is NULL, clients should subsequently call
+ * tegra_iovmm_vm_map_pages and/or tegra_iovmm_vm_insert_pfn to explicitly
+ * map the I/O virtual address to an OS-allocated page or physical address,
+ * respectively. VM operations may be called before this call returns */
+struct tegra_iovmm_area *tegra_iovmm_create_vm(
+       struct tegra_iovmm_client *client, struct tegra_iovmm_area_ops *ops,
+       unsigned long size, pgprot_t pgprot);
+
+/* called by clients to "zap" an iovmm_area, and replace all mappings
+ * in it with invalid ones, without freeing the virtual address range */
+void tegra_iovmm_zap_vm(struct tegra_iovmm_area *vm);
+
+/* after zapping a demand-loaded iovmm_area, the client should unzap it
+ * to allow the VMM device to remap the page range. */
+void tegra_iovmm_unzap_vm(struct tegra_iovmm_area *vm);
+
+/* called by clients to return an iovmm_area to the free pool for the domain */
+void tegra_iovmm_free_vm(struct tegra_iovmm_area *vm);
+
+/* called by client software to map the page-aligned I/O address vaddr to
+ * a specific physical address pfn. I/O VMA should have been created with
+ * a NULL tegra_iovmm_area_ops structure. */
+void tegra_iovmm_vm_insert_pfn(struct tegra_iovmm_area *area,
+       tegra_iovmm_addr_t vaddr, unsigned long pfn);
+
+/* called by clients to return the iovmm_area containing addr, or NULL if
+ * addr has not been allocated. caller should call tegra_iovmm_put_area when
+ * finished using the returned pointer */
+struct tegra_iovmm_area *tegra_iovmm_find_area_get(
+       struct tegra_iovmm_client *client, tegra_iovmm_addr_t addr);
+
+struct tegra_iovmm_area *tegra_iovmm_area_get(struct tegra_iovmm_area *vm);
+void tegra_iovmm_area_put(struct tegra_iovmm_area *vm);
+
+/* called by drivers to initialize a tegra_iovmm_domain structure */
+int tegra_iovmm_domain_init(struct tegra_iovmm_domain *domain,
+       struct tegra_iovmm_device *dev, tegra_iovmm_addr_t start,
+       tegra_iovmm_addr_t end);
+
+/* called by drivers to register an I/O VMM device with the system */
+int tegra_iovmm_register(struct tegra_iovmm_device *dev);
+
+/* called by drivers to remove an I/O VMM device from the system */
+int tegra_iovmm_unregister(struct tegra_iovmm_device *dev);
+
+/* called by platform suspend code to save IOVMM context */
+int tegra_iovmm_suspend(void);
+
+/* restores IOVMM context */
+void tegra_iovmm_resume(void);
+
+#else /* CONFIG_TEGRA_IOVMM */
+
+static inline struct tegra_iovmm_client *tegra_iovmm_alloc_client(
+       const char *name, const char *share_group)
+{
+       return NULL;
+}
+
+static inline size_t tegra_iovmm_get_vm_size(struct tegra_iovmm_client *client)
+{
+       return 0;
+}
+
+static inline void tegra_iovmm_free_client(struct tegra_iovmm_client *client)
+{}
+
+static inline int tegra_iovmm_client_lock(struct tegra_iovmm_client *client)
+{
+       return 0;
+}
+
+static inline int tegra_iovmm_client_trylock(struct tegra_iovmm_client *client)
+{
+       return 0;
+}
+
+static inline void tegra_iovmm_client_unlock(struct tegra_iovmm_client *client)
+{}
+
+static inline struct tegra_iovmm_area *tegra_iovmm_create_vm(
+       struct tegra_iovmm_client *client, struct tegra_iovmm_area_ops *ops,
+       unsigned long size, pgprot_t pgprot)
+{
+       return NULL;
+}
+
+static inline void tegra_iovmm_zap_vm(struct tegra_iovmm_area *vm) { }
+
+static inline void tegra_iovmm_unzap_vm(struct tegra_iovmm_area *vm) { }
+
+static inline void tegra_iovmm_free_vm(struct tegra_iovmm_area *vm) { }
+
+static inline void tegra_iovmm_vm_insert_pfn(struct tegra_iovmm_area *area,
+       tegra_iovmm_addr_t vaddr, unsigned long pfn) { }
+
+static inline struct tegra_iovmm_area *tegra_iovmm_find_area_get(
+       struct tegra_iovmm_client *client, tegra_iovmm_addr_t addr)
+{
+       return NULL;
+}
+
+static inline struct tegra_iovmm_area *tegra_iovmm_area_get(
+       struct tegra_iovmm_area *vm)
+{
+       return NULL;
+}
+
+static inline void tegra_iovmm_area_put(struct tegra_iovmm_area *vm) { }
+
+static inline int tegra_iovmm_domain_init(struct tegra_iovmm_domain *domain,
+       struct tegra_iovmm_device *dev, tegra_iovmm_addr_t start,
+       tegra_iovmm_addr_t end)
+{
+       return 0;
+}
+
+static inline int tegra_iovmm_register(struct tegra_iovmm_device *dev)
+{
+       return 0;
+}
+
+static inline int tegra_iovmm_unregister(struct tegra_iovmm_device *dev)
+{
+       return 0;
+}
+
+static inline int tegra_iovmm_suspend(void)
+{
+       return 0;
+}
+
+static inline void tegra_iovmm_resume(void) { }
+#endif /* CONFIG_TEGRA_IOVMM */
+
+
+#endif
diff --git a/arch/arm/mach-tegra/iovmm-gart.c b/arch/arm/mach-tegra/iovmm-gart.c
new file mode 100644 (file)
index 0000000..ef052e2
--- /dev/null
@@ -0,0 +1,351 @@
+/*
+ * arch/arm/mach-tegra/iovmm-gart.c
+ *
+ * Tegra I/O VMM implementation for GART devices in Tegra and Tegra 2 series
+ * systems-on-a-chip.
+ *
+ * Copyright (c) 2010, NVIDIA Corporation.
+ *
+ * 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.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
+ */
+
+#include <linux/module.h>
+#include <linux/platform_device.h>
+#include <linux/spinlock.h>
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+#include <linux/mm.h>
+#include <asm/io.h>
+#include <asm/cacheflush.h>
+
+#include <mach/iovmm.h>
+
+#if defined(CONFIG_ARCH_TEGRA_2x_SOC)
+#define GART_CONFIG            0x24
+#define GART_ENTRY_ADDR                0x28
+#define GART_ENTRY_DATA                0x2c
+#endif
+
+#define VMM_NAME "iovmm-gart"
+#define DRIVER_NAME "tegra_gart"
+
+#define GART_PAGE_SHIFT (12)
+#define GART_PAGE_MASK (~((1<<GART_PAGE_SHIFT)-1))
+
+struct gart_device {
+       void __iomem            *regs;
+       u32                     *savedata;
+       u32                     page_count; /* total remappable size */
+       tegra_iovmm_addr_t      iovmm_base; /* offset to apply to vmm_area */
+       spinlock_t              pte_lock;
+       struct tegra_iovmm_device iovmm;
+       struct tegra_iovmm_domain domain;
+       bool                    enable;
+       bool                    needs_barrier; /* emulator WAR */
+};
+
+static int gart_map(struct tegra_iovmm_device *, struct tegra_iovmm_area *);
+static void gart_unmap(struct tegra_iovmm_device *,
+       struct tegra_iovmm_area *, bool);
+static void gart_map_pfn(struct tegra_iovmm_device *,
+       struct tegra_iovmm_area *, tegra_iovmm_addr_t, unsigned long);
+static struct tegra_iovmm_domain *gart_alloc_domain(
+       struct tegra_iovmm_device *, struct tegra_iovmm_client *);
+
+static int gart_probe(struct platform_device *);
+static int gart_remove(struct platform_device *);
+static int gart_suspend(struct tegra_iovmm_device *dev);
+static void gart_resume(struct tegra_iovmm_device *dev);
+
+
+static struct tegra_iovmm_device_ops tegra_iovmm_gart_ops = {
+       .map            = gart_map,
+       .unmap          = gart_unmap,
+       .map_pfn        = gart_map_pfn,
+       .alloc_domain   = gart_alloc_domain,
+       .suspend        = gart_suspend,
+       .resume         = gart_resume,
+};
+
+static struct platform_driver tegra_iovmm_gart_drv = {
+       .probe          = gart_probe,
+       .remove         = gart_remove,
+       .driver         = {
+               .name   = DRIVER_NAME,
+       },
+};
+
+static int gart_suspend(struct tegra_iovmm_device *dev)
+{
+       struct gart_device *gart = container_of(dev, struct gart_device, iovmm);
+       unsigned int i;
+       unsigned long reg;
+
+       if (!gart)
+               return -ENODEV;
+
+       if (!gart->enable)
+               return 0;
+
+       spin_lock(&gart->pte_lock);
+       reg = gart->iovmm_base;
+       for (i=0; i<gart->page_count; i++) {
+               writel(reg, gart->regs + GART_ENTRY_ADDR);
+               gart->savedata[i] = readl(gart->regs + GART_ENTRY_DATA);
+               dmb();
+               reg += 1 << GART_PAGE_SHIFT;
+       }
+       spin_unlock(&gart->pte_lock);
+       return 0;
+}
+
+static void do_gart_setup(struct gart_device *gart, const u32 *data)
+{
+       unsigned long reg;
+       unsigned int i;
+
+       writel(1, gart->regs + GART_CONFIG);
+
+       reg = gart->iovmm_base;
+       for (i=0; i<gart->page_count; i++) {
+               writel(reg, gart->regs + GART_ENTRY_ADDR);
+               writel((data) ? data[i] : 0, gart->regs + GART_ENTRY_DATA);
+               wmb();
+               reg += 1 << GART_PAGE_SHIFT;
+       }
+       wmb();
+}
+
+static void gart_resume(struct tegra_iovmm_device *dev)
+{
+       struct gart_device *gart = container_of(dev, struct gart_device, iovmm);
+
+       if (!gart || !gart->enable || (gart->enable && !gart->savedata))
+               return;
+
+       spin_lock(&gart->pte_lock);
+       do_gart_setup(gart, gart->savedata);
+       spin_unlock(&gart->pte_lock);
+}
+
+static int gart_remove(struct platform_device *pdev)
+{
+       struct gart_device *gart = platform_get_drvdata(pdev);
+
+       if (!gart)
+               return 0;
+
+       if (gart->enable)
+               writel(0, gart->regs + GART_CONFIG);
+
+       gart->enable = 0;
+       platform_set_drvdata(pdev, NULL);
+       tegra_iovmm_unregister(&gart->iovmm);
+       if (gart->savedata)
+               vfree(gart->savedata);
+       if (gart->regs)
+               iounmap(gart->regs);
+       kfree(gart);
+       return 0;
+}
+
+static int gart_probe(struct platform_device *pdev)
+{
+       struct gart_device *gart = NULL;
+       struct resource *res, *res_remap;
+       void __iomem *gart_regs = NULL;
+       int e;
+
+       if (!pdev) {
+               pr_err(DRIVER_NAME ": platform_device required\n");
+               return -ENODEV;
+       }
+
+       if (PAGE_SHIFT != GART_PAGE_SHIFT) {
+               pr_err(DRIVER_NAME ": GART and CPU page size must match\n");
+               return -ENXIO;
+       }
+
+       /* the GART memory aperture is required */
+       res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+       res_remap = platform_get_resource(pdev, IORESOURCE_MEM, 1);
+
+       if (!res || !res_remap) {
+               pr_err(DRIVER_NAME ": GART memory aperture expected\n");
+               return -ENXIO;
+       }
+       gart = kzalloc(sizeof(*gart), GFP_KERNEL);
+       if (!gart) {
+               pr_err(DRIVER_NAME ": failed to allocate tegra_iovmm_device\n");
+               e = -ENOMEM;
+               goto fail;
+       }
+
+       gart_regs = ioremap_wc(res->start, res->end - res->start + 1);
+       if (!gart_regs) {
+               pr_err(DRIVER_NAME ": failed to remap GART registers\n");
+               e = -ENXIO;
+               goto fail;
+       }
+
+       gart->iovmm.name = VMM_NAME;
+       gart->iovmm.ops = &tegra_iovmm_gart_ops;
+       gart->iovmm.pgsize_bits = GART_PAGE_SHIFT;
+       spin_lock_init(&gart->pte_lock);
+
+       platform_set_drvdata(pdev, gart);
+
+       e = tegra_iovmm_register(&gart->iovmm);
+       if (e) goto fail;
+
+       e = tegra_iovmm_domain_init(&gart->domain, &gart->iovmm,
+               (tegra_iovmm_addr_t)res_remap->start,
+               (tegra_iovmm_addr_t)res_remap->end+1);
+       if (e) goto fail;
+
+       gart->regs = gart_regs;
+       gart->iovmm_base = (tegra_iovmm_addr_t)res_remap->start;
+       gart->page_count = res_remap->end - res_remap->start + 1;
+       gart->page_count >>= GART_PAGE_SHIFT;
+
+       gart->savedata = vmalloc(sizeof(u32)*gart->page_count);
+       if (!gart->savedata) {
+               pr_err(DRIVER_NAME ": failed to allocate context save area\n");
+               e = -ENOMEM;
+               goto fail;
+       }
+
+       spin_lock(&gart->pte_lock);
+
+       do_gart_setup(gart, NULL);
+       gart->enable = 1;
+
+       spin_unlock(&gart->pte_lock);
+       return 0;
+
+fail:
+       if (gart_regs)
+               iounmap(gart_regs);
+       if (gart && gart->savedata)
+               vfree(gart->savedata);
+       if (gart)
+               kfree(gart);
+       return e;
+}
+
+static int __devinit gart_init(void)
+{
+       return platform_driver_register(&tegra_iovmm_gart_drv);
+}
+
+static void __exit gart_exit(void)
+{
+       return platform_driver_unregister(&tegra_iovmm_gart_drv);
+}
+
+#define GART_PTE(_pfn) (0x80000000ul | ((_pfn)<<PAGE_SHIFT))
+
+
+static int gart_map(struct tegra_iovmm_device *dev,
+       struct tegra_iovmm_area *iovma)
+{
+       struct gart_device *gart = container_of(dev, struct gart_device, iovmm);
+       unsigned long gart_page, count;
+       unsigned int i;
+
+       gart_page = iovma->iovm_start;
+       count = iovma->iovm_length >> GART_PAGE_SHIFT;
+
+       for (i=0; i<count; i++) {
+               unsigned long pfn;
+
+               pfn = iovma->ops->lock_makeresident(iovma, i<<PAGE_SHIFT);
+               if (!pfn_valid(pfn))
+                       goto fail;
+
+               spin_lock(&gart->pte_lock);
+
+               writel(gart_page, gart->regs + GART_ENTRY_ADDR);
+               writel(GART_PTE(pfn), gart->regs + GART_ENTRY_DATA);
+               wmb();
+               gart_page += 1 << GART_PAGE_SHIFT;
+
+               spin_unlock(&gart->pte_lock);
+       }
+       wmb();
+       return 0;
+
+fail:
+       spin_lock(&gart->pte_lock);
+       while (i--) {
+               iovma->ops->release(iovma, i<<PAGE_SHIFT);
+               gart_page -= 1 << GART_PAGE_SHIFT;
+               writel(gart_page, gart->regs + GART_ENTRY_ADDR);
+               writel(0, gart->regs + GART_ENTRY_DATA);
+               wmb();
+       }
+       spin_unlock(&gart->pte_lock);
+       wmb();
+       return -ENOMEM;
+}
+
+static void gart_unmap(struct tegra_iovmm_device *dev,
+       struct tegra_iovmm_area *iovma, bool decommit)
+{
+       struct gart_device *gart = container_of(dev, struct gart_device, iovmm);
+       unsigned long gart_page, count;
+       unsigned int i;
+
+       count = iovma->iovm_length >> GART_PAGE_SHIFT;
+       gart_page = iovma->iovm_start;
+
+       spin_lock(&gart->pte_lock);
+       for (i=0; i<count; i++) {
+               if (iovma->ops && iovma->ops->release)
+                       iovma->ops->release(iovma, i<<PAGE_SHIFT);
+
+               writel(gart_page, gart->regs + GART_ENTRY_ADDR);
+               writel(0, gart->regs + GART_ENTRY_DATA);
+               wmb();
+               gart_page += 1 << GART_PAGE_SHIFT;
+       }
+       spin_unlock(&gart->pte_lock);
+       wmb();
+}
+
+static void gart_map_pfn(struct tegra_iovmm_device *dev,
+       struct tegra_iovmm_area *iovma, tegra_iovmm_addr_t offs,
+       unsigned long pfn)
+{
+       struct gart_device *gart = container_of(dev, struct gart_device, iovmm);
+
+       BUG_ON(!pfn_valid(pfn));
+       spin_lock(&gart->pte_lock);
+       writel(offs, gart->regs + GART_ENTRY_ADDR);
+       writel(GART_PTE(pfn), gart->regs + GART_ENTRY_DATA);
+       wmb();
+       spin_unlock(&gart->pte_lock);
+       wmb();
+}
+
+static struct tegra_iovmm_domain *gart_alloc_domain(
+       struct tegra_iovmm_device *dev, struct tegra_iovmm_client *client)
+{
+       struct gart_device *gart = container_of(dev, struct gart_device, iovmm);
+       return &gart->domain;
+}
+
+module_init(gart_init);
+module_exit(gart_exit);
diff --git a/arch/arm/mach-tegra/iovmm.c b/arch/arm/mach-tegra/iovmm.c
new file mode 100644 (file)
index 0000000..1f9e499
--- /dev/null
@@ -0,0 +1,775 @@
+/*
+ * arch/arm/mach-tegra/iovmm.c
+ *
+ * Tegra I/O VM manager
+ *
+ * Copyright (c) 2010, NVIDIA Corporation.
+ *
+ * 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.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
+ */
+
+#include <linux/kernel.h>
+#include <linux/spinlock.h>
+#include <linux/proc_fs.h>
+#include <linux/sched.h>
+#include <linux/string.h>
+#include <linux/slab.h>
+
+#include <mach/iovmm.h>
+
+/* after the best-fit block is located, the remaining pages not needed for
+ * the allocation will be split into a new free block if the number of
+ * remaining pages is >= MIN_SPLIT_PAGE.
+ */
+#define MIN_SPLIT_PAGE (4)
+#define MIN_SPLIT_BYTES(_d) (MIN_SPLIT_PAGE<<(_d)->dev->pgsize_bits)
+
+#define iovmm_start(_b) ((_b)->vm_area.iovm_start)
+#define iovmm_length(_b) ((_b)->vm_area.iovm_length)
+#define iovmm_end(_b) (iovmm_start(_b) + iovmm_length(_b))
+
+/* flags for the block */
+#define BK_free                0 /* indicates free mappings */
+#define BK_map_dirty   1 /* used by demand-loaded mappings */
+
+/* flags for the client */
+#define CL_locked      0
+
+/* flags for the domain */
+#define DM_map_dirty   0
+
+struct tegra_iovmm_block {
+       struct tegra_iovmm_area vm_area;
+       atomic_t                ref;
+       unsigned long           flags;
+       unsigned long           poison;
+       struct rb_node          free_node;
+       struct rb_node          all_node;
+};
+
+struct iovmm_share_group {
+       const char                      *name;
+       struct tegra_iovmm_domain       *domain;
+       struct list_head                client_list;
+       struct list_head                group_list;
+       spinlock_t                      lock;
+};
+
+static LIST_HEAD(iovmm_devices);
+static LIST_HEAD(iovmm_groups);
+static DEFINE_MUTEX(iovmm_list_lock);
+static struct kmem_cache *iovmm_cache;
+
+static tegra_iovmm_addr_t iovmm_align_up(struct tegra_iovmm_device *dev,
+       tegra_iovmm_addr_t addr)
+{
+       addr += (1<<dev->pgsize_bits);
+       addr--;
+       addr &= ~((1<<dev->pgsize_bits)-1);
+       return addr;
+}
+
+static tegra_iovmm_addr_t iovmm_align_down(struct tegra_iovmm_device *dev,
+       tegra_iovmm_addr_t addr)
+{
+       addr &= ~((1<<dev->pgsize_bits)-1);
+       return addr;
+}
+
+#define iovmprint(fmt, arg...) snprintf(page+len, count-len, fmt, ## arg)
+
+static void tegra_iovmm_block_stats(struct tegra_iovmm_domain *domain,
+       unsigned int *num_blocks, unsigned int *num_free,
+       tegra_iovmm_addr_t *total, tegra_iovmm_addr_t *total_free,
+       tegra_iovmm_addr_t *max_free)
+{
+       struct rb_node *n;
+       struct tegra_iovmm_block *b;
+
+       *num_blocks = 0;
+       *num_free = 0;
+       *total = (tegra_iovmm_addr_t)0;
+       *total_free = (tegra_iovmm_addr_t)0;
+       *max_free = (tegra_iovmm_addr_t)0;
+
+       spin_lock(&domain->block_lock);
+       n = rb_first(&domain->all_blocks);
+       while (n) {
+               b = rb_entry(n, struct tegra_iovmm_block, all_node);
+               n = rb_next(n);
+               (*num_blocks)++;
+               (*total) += iovmm_length(b);
+               if (test_bit(BK_free, &b->flags)) {
+                       (*num_free)++;
+                       (*total_free) += iovmm_length(b);
+                       (*max_free) = max_t(tegra_iovmm_addr_t,
+                               (*max_free), iovmm_length(b));
+               }
+       }
+       spin_unlock(&domain->block_lock);
+}
+
+static int tegra_iovmm_read_proc(char *page, char **start, off_t off,
+       int count, int *eof, void *data)
+{
+       struct iovmm_share_group *grp;
+       tegra_iovmm_addr_t max_free, total_free, total;
+       unsigned int num, num_free;
+
+       int len = 0;
+
+       mutex_lock(&iovmm_list_lock);
+       len += iovmprint("\ngroups\n");
+       if (list_empty(&iovmm_groups))
+               len += iovmprint("\t<empty>\n");
+       else {
+               list_for_each_entry(grp, &iovmm_groups, group_list) {
+                       len += iovmprint("\t%s (device: %s)\n",
+                               (grp->name) ? grp->name : "<unnamed>",
+                               grp->domain->dev->name);
+                       tegra_iovmm_block_stats(grp->domain, &num,
+                               &num_free, &total, &total_free, &max_free);
+                       total >>= 10;
+                       total_free >>= 10;
+                       max_free >>= 10;
+                       len += iovmprint("\t\tsize: %uKiB free: %uKiB "
+                               "largest: %uKiB (%u free / %u total blocks)\n",
+                               total, total_free, max_free, num_free, num);
+               }
+       }
+       mutex_unlock(&iovmm_list_lock);
+
+       *eof = 1;
+       return len;
+}
+
+static void iovmm_block_put(struct tegra_iovmm_block *b)
+{
+       BUG_ON(b->poison);
+       BUG_ON(atomic_read(&b->ref)==0);
+       if (!atomic_dec_return(&b->ref)) {
+               b->poison = 0xa5a5a5a5;
+               kmem_cache_free(iovmm_cache, b);
+       }
+}
+
+static void iovmm_free_block(struct tegra_iovmm_domain *domain,
+       struct tegra_iovmm_block *block)
+{
+       struct tegra_iovmm_block *pred = NULL; /* address-order predecessor */
+       struct tegra_iovmm_block *succ = NULL; /* address-order successor */
+       struct rb_node **p;
+       struct rb_node *parent = NULL, *temp;
+       int pred_free = 0, succ_free = 0;
+
+       iovmm_block_put(block);
+
+       spin_lock(&domain->block_lock);
+       temp = rb_prev(&block->all_node);
+       if (temp)
+               pred = rb_entry(temp, struct tegra_iovmm_block, all_node);
+       temp = rb_next(&block->all_node);
+       if (temp)
+               succ = rb_entry(temp, struct tegra_iovmm_block, all_node);
+
+       if (pred) pred_free = test_bit(BK_free, &pred->flags);
+       if (succ) succ_free = test_bit(BK_free, &succ->flags);
+
+       if (pred_free && succ_free) {
+               iovmm_length(pred) += iovmm_length(block);
+               iovmm_length(pred) += iovmm_length(succ);
+               rb_erase(&block->all_node, &domain->all_blocks);
+               rb_erase(&succ->all_node, &domain->all_blocks);
+               rb_erase(&succ->free_node, &domain->free_blocks);
+               rb_erase(&pred->free_node, &domain->free_blocks);
+               iovmm_block_put(block);
+               iovmm_block_put(succ);
+               block = pred;
+       } else if (pred_free) {
+               iovmm_length(pred) += iovmm_length(block);
+               rb_erase(&block->all_node, &domain->all_blocks);
+               rb_erase(&pred->free_node, &domain->free_blocks);
+               iovmm_block_put(block);
+               block = pred;
+       } else if (succ_free) {
+               iovmm_length(block) += iovmm_length(succ);
+               rb_erase(&succ->all_node, &domain->all_blocks);
+               rb_erase(&succ->free_node, &domain->free_blocks);
+               iovmm_block_put(succ);
+       }
+
+       p = &domain->free_blocks.rb_node;
+       while (*p) {
+               struct tegra_iovmm_block *b;
+               parent = *p;
+               b = rb_entry(parent, struct tegra_iovmm_block, free_node);
+               if (iovmm_length(block) >= iovmm_length(b))
+                       p = &parent->rb_right;
+               else
+                       p = &parent->rb_left;
+       }
+       rb_link_node(&block->free_node, parent, p);
+       rb_insert_color(&block->free_node, &domain->free_blocks);
+       set_bit(BK_free, &block->flags);
+       spin_unlock(&domain->block_lock);
+}
+
+/* if the best-fit block is larger than the requested size, a remainder
+ * block will be created and inserted into the free list in its place.
+ * since all free blocks are stored in two trees the new block needs to be
+ * linked into both. */
+static void iovmm_split_free_block(struct tegra_iovmm_domain *domain,
+       struct tegra_iovmm_block *block, unsigned long size)
+{
+       struct rb_node **p;
+       struct rb_node *parent = NULL;
+       struct tegra_iovmm_block *rem;
+       struct tegra_iovmm_block *b;
+
+       rem = kmem_cache_zalloc(iovmm_cache, GFP_KERNEL);
+       if (!rem) return;
+
+       spin_lock(&domain->block_lock);
+       p = &domain->free_blocks.rb_node;
+
+       iovmm_start(rem) = iovmm_start(block) + size;
+       iovmm_length(rem) = iovmm_length(block) - size;
+       atomic_set(&rem->ref, 1);
+       iovmm_length(block) = size;
+
+       while (*p) {
+               parent = *p;
+               b = rb_entry(parent, struct tegra_iovmm_block, free_node);
+               if (iovmm_length(rem) >= iovmm_length(b))
+                       p = &parent->rb_right;
+               else
+                       p = &parent->rb_left;
+       }
+       set_bit(BK_free, &rem->flags);
+       rb_link_node(&rem->free_node, parent, p);
+       rb_insert_color(&rem->free_node, &domain->free_blocks);
+
+       p = &domain->all_blocks.rb_node;
+       parent = NULL;
+       while (*p) {
+               parent = *p;
+               b = rb_entry(parent, struct tegra_iovmm_block, all_node);
+               if (iovmm_start(rem) >= iovmm_start(b))
+                       p = &parent->rb_right;
+               else
+                       p = &parent->rb_left;
+       }
+       rb_link_node(&rem->all_node, parent, p);
+       rb_insert_color(&rem->all_node, &domain->all_blocks);
+}
+
+static struct tegra_iovmm_block *iovmm_alloc_block(
+       struct tegra_iovmm_domain *domain, unsigned long size)
+{
+       struct rb_node *n;
+       struct tegra_iovmm_block *b, *best;
+        static int splitting = 0;
+
+       BUG_ON(!size);
+       size = iovmm_align_up(domain->dev, size);
+        for (;;) {
+               spin_lock(&domain->block_lock);
+               if (!splitting)
+                       break;
+               spin_unlock(&domain->block_lock);
+               schedule();
+       }
+       n = domain->free_blocks.rb_node;
+       best = NULL;
+       while (n) {
+               b = rb_entry(n, struct tegra_iovmm_block, free_node);
+               if (iovmm_length(b) < size) n = n->rb_right;
+               else if (iovmm_length(b) == size) {
+                       best = b;
+                       break;
+               } else {
+                       best = b;
+                       n = n->rb_left;
+               }
+       }
+       if (!best) {
+               spin_unlock(&domain->block_lock);
+               return NULL;
+       }
+       rb_erase(&best->free_node, &domain->free_blocks);
+       clear_bit(BK_free, &best->flags);
+       atomic_inc(&best->ref);
+       if (iovmm_length(best) >= size+MIN_SPLIT_BYTES(domain)) {
+               splitting = 1;
+               spin_unlock(&domain->block_lock);
+               iovmm_split_free_block(domain, best, size);
+               splitting = 0;
+       }
+
+       spin_unlock(&domain->block_lock);
+
+       return best;
+}
+
+int tegra_iovmm_domain_init(struct tegra_iovmm_domain *domain,
+       struct tegra_iovmm_device *dev, tegra_iovmm_addr_t start,
+       tegra_iovmm_addr_t end)
+{
+       struct tegra_iovmm_block *b;
+
+       b = kmem_cache_zalloc(iovmm_cache, GFP_KERNEL);
+       if (!b) return -ENOMEM;
+
+       domain->dev = dev;
+       atomic_set(&domain->clients, 0);
+       atomic_set(&domain->locks, 0);
+       atomic_set(&b->ref, 1);
+       spin_lock_init(&domain->block_lock);
+       init_rwsem(&domain->map_lock);
+       init_waitqueue_head(&domain->delay_lock);
+       iovmm_start(b) = iovmm_align_up(dev, start);
+       iovmm_length(b) = iovmm_align_down(dev, end) - iovmm_start(b);
+       set_bit(BK_free, &b->flags);
+       rb_link_node(&b->free_node, NULL, &domain->free_blocks.rb_node);
+       rb_insert_color(&b->free_node, &domain->free_blocks);
+       rb_link_node(&b->all_node, NULL, &domain->all_blocks.rb_node);
+       rb_insert_color(&b->all_node, &domain->all_blocks);
+       return 0;
+}
+
+struct tegra_iovmm_area *tegra_iovmm_create_vm(
+       struct tegra_iovmm_client *client, struct tegra_iovmm_area_ops *ops,
+       unsigned long size, pgprot_t pgprot)
+{
+       struct tegra_iovmm_block *b;
+       struct tegra_iovmm_device *dev;
+
+       if (!client) return NULL;
+
+       dev = client->domain->dev;
+
+       b = iovmm_alloc_block(client->domain, size);
+       if (!b) return NULL;
+
+       b->vm_area.domain = client->domain;
+       b->vm_area.pgprot = pgprot;
+       b->vm_area.ops = ops;
+
+       down_read(&b->vm_area.domain->map_lock);
+       if (ops && !test_bit(CL_locked, &client->flags)) {
+               set_bit(BK_map_dirty, &b->flags);
+               set_bit(DM_map_dirty, &client->domain->flags);
+       } else if (ops) {
+               if (dev->ops->map(dev, &b->vm_area))
+                       pr_err("%s failed to map locked domain\n", __func__);
+       }
+       up_read(&b->vm_area.domain->map_lock);
+
+       return &b->vm_area;
+}
+
+void tegra_iovmm_vm_insert_pfn(struct tegra_iovmm_area *area,
+       tegra_iovmm_addr_t vaddr, unsigned long pfn)
+{
+       struct tegra_iovmm_device *dev = area->domain->dev;
+       BUG_ON(vaddr & ((1<<dev->pgsize_bits)-1));
+       BUG_ON(vaddr >= area->iovm_start + area->iovm_length);
+       BUG_ON(vaddr < area->iovm_start);
+       BUG_ON(area->ops);
+
+       dev->ops->map_pfn(dev, area, vaddr, pfn);
+}
+
+void tegra_iovmm_zap_vm(struct tegra_iovmm_area *vm)
+{
+       struct tegra_iovmm_block *b;
+       struct tegra_iovmm_device *dev;
+
+       b = container_of(vm, struct tegra_iovmm_block, vm_area);
+       dev = vm->domain->dev;
+       /* if the vm area mapping was deferred, don't unmap it since
+        * the memory for the page tables it uses may not be allocated */
+       down_read(&vm->domain->map_lock);
+       if (!test_and_clear_bit(BK_map_dirty, &b->flags))
+               dev->ops->unmap(dev, vm, false);
+       up_read(&vm->domain->map_lock);
+}
+
+void tegra_iovmm_unzap_vm(struct tegra_iovmm_area *vm)
+{
+       struct tegra_iovmm_block *b;
+       struct tegra_iovmm_device *dev;
+
+       b = container_of(vm, struct tegra_iovmm_block, vm_area);
+       dev = vm->domain->dev;
+       if (!vm->ops) return;
+
+       down_read(&vm->domain->map_lock);
+       if (vm->ops) {
+               if (atomic_read(&vm->domain->locks))
+                       dev->ops->map(dev, vm);
+               else {
+                       set_bit(BK_map_dirty, &b->flags);
+                       set_bit(DM_map_dirty, &vm->domain->flags);
+               }
+        }
+       up_read(&vm->domain->map_lock);
+}
+
+void tegra_iovmm_free_vm(struct tegra_iovmm_area *vm)
+{
+       struct tegra_iovmm_block *b;
+       struct tegra_iovmm_device *dev;
+       struct tegra_iovmm_domain *domain;
+
+       if (!vm) return;
+
+       b = container_of(vm, struct tegra_iovmm_block, vm_area);
+       domain = vm->domain;
+       dev = vm->domain->dev;
+       down_read(&domain->map_lock);
+       if (!test_and_clear_bit(BK_map_dirty, &b->flags))
+               dev->ops->unmap(dev, vm, true);
+       iovmm_free_block(domain, b);
+       up_read(&domain->map_lock);
+}
+
+struct tegra_iovmm_area *tegra_iovmm_area_get(struct tegra_iovmm_area *vm)
+{
+       struct tegra_iovmm_block *b;
+
+       BUG_ON(!vm);
+       b = container_of(vm, struct tegra_iovmm_block, vm_area);
+
+       atomic_inc(&b->ref);
+       return &b->vm_area;
+}
+
+void tegra_iovmm_area_put(struct tegra_iovmm_area *vm)
+{
+       struct tegra_iovmm_block *b;
+       BUG_ON(!vm);
+       b = container_of(vm, struct tegra_iovmm_block, vm_area);
+       iovmm_block_put(b);
+}
+
+struct tegra_iovmm_area *tegra_iovmm_find_area_get(
+       struct tegra_iovmm_client *client, tegra_iovmm_addr_t addr)
+{
+       struct rb_node *n;
+       struct tegra_iovmm_block *b = NULL;
+
+       if (!client) return NULL;
+
+       spin_lock(&client->domain->block_lock);
+       n = client->domain->all_blocks.rb_node;
+
+       while (n) {
+               b = rb_entry(n, struct tegra_iovmm_block, all_node);
+               if ((iovmm_start(b) <= addr) && (iovmm_end(b) >= addr)) {
+                       if (test_bit(BK_free, &b->flags)) b = NULL;
+                       break;
+               }
+               if (addr > iovmm_start(b))
+                       n = n->rb_right;
+               else
+                       n = n->rb_left;
+               b = NULL;
+       }
+       if (b) atomic_inc(&b->ref);
+       spin_unlock(&client->domain->block_lock);
+       if (!b) return NULL;
+       return &b->vm_area;
+}
+
+static int _iovmm_client_lock(struct tegra_iovmm_client *client)
+{
+       struct tegra_iovmm_device *dev;
+       struct tegra_iovmm_domain *domain;
+       int v;
+
+       if (unlikely(!client)) return -ENODEV;
+       if (unlikely(test_bit(CL_locked, &client->flags))) {
+               pr_err("attempting to relock client %s\n", client->name);
+               return 0;
+       }
+
+       domain = client->domain;
+       dev = domain->dev;
+       down_write(&domain->map_lock);
+       v = atomic_inc_return(&domain->locks);
+       /* if the device doesn't export the lock_domain function, the device
+        * must guarantee that any valid domain will be locked. */
+       if (v==1 && dev->ops->lock_domain) {
+               if (dev->ops->lock_domain(dev, domain)) {
+                       atomic_dec(&domain->locks);
+                       up_write(&domain->map_lock);
+                       return -EAGAIN;
+               }
+       }
+       if (test_and_clear_bit(DM_map_dirty, &domain->flags)) {
+               struct rb_node *n;
+               struct tegra_iovmm_block *b;
+
+               spin_lock(&domain->block_lock);
+               n = rb_first(&domain->all_blocks);
+               while (n) {
+                       b = rb_entry(n, struct tegra_iovmm_block, all_node);
+                       n = rb_next(n);
+                       if (test_bit(BK_free, &b->flags))
+                               continue;
+
+                       if (test_and_clear_bit(BK_map_dirty, &b->flags)) {
+                               if (!b->vm_area.ops) {
+                                       pr_err("%s: vm_area ops must exist for lazy maps\n", __func__);
+                                       continue;
+                               }
+                               dev->ops->map(dev, &b->vm_area);
+                       }
+               }
+       }
+       set_bit(CL_locked, &client->flags);
+       up_write(&domain->map_lock);
+       return 0;
+}
+
+int tegra_iovmm_client_trylock(struct tegra_iovmm_client *client)
+{
+       return _iovmm_client_lock(client);
+}
+
+int tegra_iovmm_client_lock(struct tegra_iovmm_client *client)
+{
+       int ret;
+
+       if (!client) return -ENODEV;
+
+       ret = wait_event_interruptible(client->domain->delay_lock,
+               _iovmm_client_lock(client)!=-EAGAIN);
+
+       if (ret==-ERESTARTSYS) return -EINTR;
+
+       return ret;
+}
+
+void tegra_iovmm_client_unlock(struct tegra_iovmm_client *client)
+{
+       struct tegra_iovmm_device *dev;
+       struct tegra_iovmm_domain *domain;
+       int do_wake = 0;
+
+       if (!client) return;
+
+       if (!test_and_clear_bit(CL_locked, &client->flags)) {
+               pr_err("unlocking unlocked client %s\n", client->name);
+               return;
+       }
+
+       domain = client->domain;
+       dev = domain->dev;
+        down_write(&domain->map_lock);
+       if (!atomic_dec_return(&client->domain->locks)) {
+               if (dev->ops->unlock_domain)
+                       dev->ops->unlock_domain(dev, domain);
+               do_wake = 1;
+       }
+       up_write(&domain->map_lock);
+       if (do_wake) wake_up(&domain->delay_lock);
+}
+
+size_t tegra_iovmm_get_vm_size(struct tegra_iovmm_client *client)
+{
+       struct tegra_iovmm_domain *domain;
+       struct rb_node *n;
+       struct tegra_iovmm_block *b;
+       size_t size = 0;
+
+       if (!client) return 0;
+
+       domain = client->domain;
+
+       spin_lock(&domain->block_lock);
+       n = rb_first(&domain->all_blocks);
+       while (n) {
+               b = rb_entry(n, struct tegra_iovmm_block, all_node);
+               n = rb_next(n);
+               size += iovmm_length(b);
+       }
+       spin_unlock(&domain->block_lock);
+
+       return size;
+}
+
+void tegra_iovmm_free_client(struct tegra_iovmm_client *client)
+{
+       struct tegra_iovmm_device *dev;
+       if (!client) return;
+
+       BUG_ON(!client->domain || !client->domain->dev);
+
+       dev = client->domain->dev;
+
+       if (test_and_clear_bit(CL_locked, &client->flags)) {
+               pr_err("freeing locked client %s\n", client->name);
+               if (!atomic_dec_return(&client->domain->locks)) {
+                       down_write(&client->domain->map_lock);
+                       if (dev->ops->unlock_domain)
+                               dev->ops->unlock_domain(dev, client->domain);
+                       up_write(&client->domain->map_lock);
+                       wake_up(&client->domain->delay_lock);
+               }
+       }
+       mutex_lock(&iovmm_list_lock);
+       if (!atomic_dec_return(&client->domain->clients))
+               if (dev->ops->free_domain)
+                       dev->ops->free_domain(dev, client->domain);
+       list_del(&client->list);
+       if (list_empty(&client->group->client_list)) {
+               list_del(&client->group->group_list);
+               if (client->group->name) kfree(client->group->name);
+               kfree(client->group);
+       }
+       kfree(client->name);
+       kfree(client);
+       mutex_unlock(&iovmm_list_lock);
+}
+
+struct tegra_iovmm_client *tegra_iovmm_alloc_client(const char *name,
+       const char *share_group)
+{
+       struct tegra_iovmm_client *c = kzalloc(sizeof(*c), GFP_KERNEL);
+       struct iovmm_share_group *grp = NULL;
+       struct tegra_iovmm_device *dev;
+
+       if (!c) return NULL;
+       c->name = kstrdup(name, GFP_KERNEL);
+       if (!c->name) goto fail;
+
+       mutex_lock(&iovmm_list_lock);
+       if (share_group) {
+               list_for_each_entry(grp, &iovmm_groups, group_list) {
+                       if (grp->name && !strcmp(grp->name, share_group))
+                               break;
+               }
+       }
+       if (!grp || strcmp(grp->name, share_group)) {
+               grp = kzalloc(sizeof(*grp), GFP_KERNEL);
+               if (!grp) goto fail_lock;
+               grp->name = (share_group) ? kstrdup(share_group, GFP_KERNEL) : NULL;
+               if (share_group && !grp->name) {
+                       kfree(grp);
+                       goto fail_lock;
+               }
+               list_for_each_entry(dev, &iovmm_devices, list) {
+                       grp->domain = dev->ops->alloc_domain(dev, c);
+                       if (grp->domain) break;
+               }
+               if (!grp->domain) {
+                       pr_err("%s: alloc_domain failed for %s\n",
+                               __func__, c->name);
+                       dump_stack();
+                       if (grp->name) kfree(grp->name);
+                       kfree(grp);
+                       grp = NULL;
+                       goto fail_lock;
+                }
+               spin_lock_init(&grp->lock);
+               INIT_LIST_HEAD(&grp->client_list);
+               list_add_tail(&grp->group_list, &iovmm_groups);
+       }
+
+       atomic_inc(&grp->domain->clients);
+       c->group = grp;
+       c->domain = grp->domain;
+       spin_lock(&grp->lock);
+       list_add_tail(&c->list, &grp->client_list);
+       spin_unlock(&grp->lock);
+       mutex_unlock(&iovmm_list_lock);
+       return c;
+
+fail_lock:
+       mutex_unlock(&iovmm_list_lock);
+fail:
+       if (c) {
+               if (c->name) kfree(c->name);
+               kfree(c);
+       }
+       return NULL;
+}
+
+int tegra_iovmm_register(struct tegra_iovmm_device *dev)
+{
+       BUG_ON(!dev);
+       mutex_lock(&iovmm_list_lock);
+       if (list_empty(&iovmm_devices)) {
+               iovmm_cache = KMEM_CACHE(tegra_iovmm_block, 0);
+               if (!iovmm_cache) {
+                       pr_err("%s: failed to make kmem cache\n", __func__);
+                       mutex_unlock(&iovmm_list_lock);
+                       return -ENOMEM;
+               }
+               create_proc_read_entry("iovmminfo", S_IRUGO, NULL,
+                       tegra_iovmm_read_proc, NULL);
+       }
+       list_add_tail(&dev->list, &iovmm_devices);
+       mutex_unlock(&iovmm_list_lock);
+       printk("%s: added %s\n", __func__, dev->name);
+       return 0;
+}
+
+int tegra_iovmm_suspend(void)
+{
+       int rc = 0;
+       struct tegra_iovmm_device *dev;
+
+       mutex_lock(&iovmm_list_lock);
+       list_for_each_entry(dev, &iovmm_devices, list) {
+
+               if (!dev->ops->suspend)
+                       continue;
+
+               rc = dev->ops->suspend(dev);
+               if (rc) {
+                       pr_err("%s: %s suspend returned %d\n",
+                              __func__, dev->name, rc);
+                       mutex_unlock(&iovmm_list_lock);
+                       return rc;
+               }
+       }
+       mutex_unlock(&iovmm_list_lock);
+       return 0;       
+}
+
+void tegra_iovmm_resume(void)
+{
+       struct tegra_iovmm_device *dev;
+
+       mutex_lock(&iovmm_list_lock);
+
+       list_for_each_entry(dev, &iovmm_devices, list) {
+               if (dev->ops->resume)
+                       dev->ops->resume(dev);
+       }
+
+       mutex_unlock(&iovmm_list_lock);
+}
+
+int tegra_iovmm_unregister(struct tegra_iovmm_device *dev)
+{
+       mutex_lock(&iovmm_list_lock);
+       list_del(&dev->list);
+       mutex_unlock(&iovmm_list_lock);
+       return 0;
+}