openwrt/staging/blogic.git
10 years agoion: use alloc_pages in system contig heap
Colin Cross [Sat, 14 Dec 2013 03:26:27 +0000 (19:26 -0800)]
ion: use alloc_pages in system contig heap

There is no reason to use kzalloc, just call alloc_pages directly.
Change the GFP from GFP_KERNEL to include __GFP_HIGH, to allow it
to return contiguous pages from highmem.  virt_to_* functions
aren't valid on highmem pages, so store the struct page * in an
sg_table in buffer->priv_virt like most other heaps, and replace
virt_to_* with page_to_*.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: allow cached mappings of chunk and system heap buffers
Colin Cross [Sat, 14 Dec 2013 03:26:26 +0000 (19:26 -0800)]
ion: allow cached mappings of chunk and system heap buffers

Now that ion_vm_fault uses vm_insert_pfn instead of vm_insert_page
cached buffers can be supported in any heap.  Remove the checks
in the chunk and system heaps.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: remove ion_heap_alloc_pages
Colin Cross [Sat, 14 Dec 2013 03:26:25 +0000 (19:26 -0800)]
ion: remove ion_heap_alloc_pages

Now that ion_vm_fault doesn't need a struct page with a nonzero
refcount, there is no need allocate heap memory for cached pages using
split_page.  Remove the ion_heap_alloc_pages and ion_heap_free_pages
helpers in favor of direct calls to alloc_pages and __free_pages,
and remove the special handling in the system heap.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: use vm_insert_pfn for faulted pages
Colin Cross [Sat, 14 Dec 2013 03:26:24 +0000 (19:26 -0800)]
ion: use vm_insert_pfn for faulted pages

Most ion userspace mappings are created with remap_pfn_range.  Use
vm_insert_pfn instead of vm_insert_page to make faulted cached
mappings look more like uncached mappings.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: check return value from remap_pfn_range
Colin Cross [Sat, 14 Dec 2013 03:26:23 +0000 (19:26 -0800)]
ion: check return value from remap_pfn_range

Check the return value of remap_pfn_range and return an error if
it fails.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: free low memory from page pools first
Colin Cross [Sat, 14 Dec 2013 03:26:22 +0000 (19:26 -0800)]
ion: free low memory from page pools first

When the shrinkers are called with GFP_HIGH free low memory first,
it is more important to have free than high memory.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: optimize ion_heap_buffer_zero
Colin Cross [Sat, 14 Dec 2013 03:26:21 +0000 (19:26 -0800)]
ion: optimize ion_heap_buffer_zero

ion_heap_buffer_zero can spend a long time in unmap_kernel_range
if it has to broadcast a tlb flush to every cpu for every page.
Modify it to batch pages into a larger region to clear using a
single mapping.  This may cause the mapping size to change if
the buffer size is not a multiple of the mapping size, so
switch to allocating the address space for each chunk.  This
allows us to use vm_map_ram to handle the allocation and mapping
together.

The number of pages to zero using a single mapping is set to 32
to hit the fastpath in vm_map_ram.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: add alignment check to carveout heap
Colin Cross [Sat, 14 Dec 2013 03:26:20 +0000 (19:26 -0800)]
ion: add alignment check to carveout heap

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: drop dependency on ARM
Colin Cross [Sat, 14 Dec 2013 03:26:19 +0000 (19:26 -0800)]
ion: drop dependency on ARM

Ion will compile and run on other platforms now, remove the
dependency on ARM.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: Fix two small issues in system_heap allocation
John Stultz [Sat, 14 Dec 2013 03:26:18 +0000 (19:26 -0800)]
ion: Fix two small issues in system_heap allocation

In testing ion system heap allocations, I ran across two issues:

1) Not k*z*allocing the sg table. This can cause trouble if
we end up trying call sg_alloc_table() with too many entries,
then sg_alloc_table() internally fails and tries to free what it
thinks is internal table structure, which causes bad pointer
traversals.

2) The second list_for_each_entry probably should be _safe,
since I was seeing  strange lock warnings and oopses on occasion.
This seems to resolve it, but could use some extra checking.

Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: fix use-after-free in ion_heap_freelist_drain
Mitchel Humpherys [Sat, 14 Dec 2013 03:26:17 +0000 (19:26 -0800)]
gpu: ion: fix use-after-free in ion_heap_freelist_drain

The `buffer' variable is being used after being freed. Fix this.

Signed-off-by: Mitchel Humpherys <mitchelh@codeaurora.org>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: clean up ioctls
Colin Cross [Sat, 14 Dec 2013 03:26:16 +0000 (19:26 -0800)]
ion: clean up ioctls

Convert the ion ioctls to use _IOW instead of _IOWR where
appropriate, and factor out the copy_from_user and copy_to_user
based on the _IOC_DIR bits.  For the existing incorrect ioctls,
add a function to wrap _IOC_DIR to return the corrected value.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: remove unnecessary function from system heap
Colin Cross [Sat, 14 Dec 2013 03:26:15 +0000 (19:26 -0800)]
gpu: ion: remove unnecessary function from system heap

ion_system_contig_heap buffers have an sglist, just call
ion_heap_map_user to map it.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: fix printk warnings
Colin Cross [Sat, 14 Dec 2013 03:26:14 +0000 (19:26 -0800)]
ion: fix printk warnings

Use %z for size_t and %pa for dma_addr_t to avoid warnings in printks.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: don't use phys_to_page or __phys_to_pfn
Colin Cross [Sat, 14 Dec 2013 03:26:13 +0000 (19:26 -0800)]
ion: don't use phys_to_page or __phys_to_pfn

phys_to_page and __phys_to_pfn don't exist on all platforms.
Use a combination of pfn_to_page, PFN_DOWN, page_to_pfn, and
virt_to_page to get the same results.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: don't use __arm_ioremap to map pages
Colin Cross [Fri, 13 Dec 2013 22:25:06 +0000 (14:25 -0800)]
ion: don't use __arm_ioremap to map pages

ion_heap_map_kernel already implements mapping a scatterlist of
pages into the kernel, and all heaps are required to have struct
pages associated with them, so delete the functions that use
__arm_ioremap and use ion_heap_map_kernel instead.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: update idr to avoid deprecated apis
Colin Cross [Fri, 13 Dec 2013 22:25:05 +0000 (14:25 -0800)]
ion: update idr to avoid deprecated apis

Use idr_alloc instead if idr_pre_get/idr_get_new_above, and
remove idr_remove_all.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: add test device for unit tests to interact with dma_bufs
Colin Cross [Fri, 13 Dec 2013 22:25:04 +0000 (14:25 -0800)]
ion: add test device for unit tests to interact with dma_bufs

Add a /dev/ion-test device that will be created if CONFIG_ION_TEST
is set.  The device accepts a dma_buf fd and allows reading and
writing to the backing memory using DMA-like apis or kernel mapping
apis.  Can be used to test the dma_buf mapping ops, including
the ion implementations, from userspace.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: check invalid values in ion_system_heap
Colin Cross [Fri, 13 Dec 2013 22:25:03 +0000 (14:25 -0800)]
ion: check invalid values in ion_system_heap

ion_system_heap can only satisfy page alignment, and
ion_system_contig_heap can only satisify alignment to the
allocation size.  Neither can support faulting user mappings
because they use slab pages.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: convert sg_dma_len(sg) to sg->length
Colin Cross [Fri, 13 Dec 2013 22:25:02 +0000 (14:25 -0800)]
ion: convert sg_dma_len(sg) to sg->length

ion is always dealing with the allocation and not the mapping,
so it should always be using sg->length and not sg->dma_length.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: fix dma APIs
Colin Cross [Fri, 13 Dec 2013 22:25:01 +0000 (14:25 -0800)]
ion: fix dma APIs

__dma_page_cpu_to_dev is a private ARM api that is not available
on 3.10 and was never available on other architectures.  We can
get the same behavior by calling dma_sync_sg_for_device with a
scatterlist containing a single page.  It's still not quite a
kosher use of the dma apis, we still conflate physical addresses
with bus addresses, but it should at least compile on all
platforms, and work on any platform that doesn't have a physical
to bus address translation.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: fix crash when alloc len is -1
Colin Cross [Fri, 13 Dec 2013 22:25:00 +0000 (14:25 -0800)]
ion: fix crash when alloc len is -1

If userspace passes a length between -4095 and -1 to allocate it
will pass the len != 0 check, but when len is page aligned it will
be 0.  Check len after page aligning.

Drop the warning as well, userspace shouldn't be able to trigger
a warning in the kernel.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: hold reference to handle after ion_uhandle_get
Colin Cross [Fri, 13 Dec 2013 22:24:59 +0000 (14:24 -0800)]
ion: hold reference to handle after ion_uhandle_get

commit 1262ab1846cf76f7549c66ef709120dbfbe6d49f (ion: replace
userspace handle cookies with idr) broke the locking in ion.
The ION_IOC_FREE and ION_IOC_MAP ioctls were relying on
ion_handle_validate to detect the case where a call raced
with another ION_IOC_FREE which may have freed the struct
ion_handle.

Rename ion_uhandle_get to ion_handle_get_by_id, and have it
take the client lock and return with an extra reference to
the handle.  Make each caller put its reference once it
is done with the handle.

Also modify users of ion_handle_validate to continue to hold
the client lock after calling ion_handle_validate until
they are done with the handle, and warn if ion_handle_validate
is called without the client lock held.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: Fix compat support to use proper compat ioctl numbers
John Stultz [Fri, 13 Dec 2013 22:24:58 +0000 (14:24 -0800)]
ion: Fix compat support to use proper compat ioctl numbers

The compat support added to ion didn't provide compat ioctl numbers
(who's value depends on the compat structure size). So 32bit
applications would get an error when trying to make ioctl calls.

This patch adds the needed COMPAT_ macros and uses them in the
compat_ion_ioctl, translating them to their non-compat cmd when
calling the normal ioctl call.

Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: move userspace api into uapi/ion.h
Colin Cross [Fri, 13 Dec 2013 22:24:57 +0000 (14:24 -0800)]
ion: move userspace api into uapi/ion.h

Split the userspace api out of drivers/staging/android/ion/ion.h
into drivers/staging/android/uapi/ion.h

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: delete ion_system_mapper.c
Colin Cross [Fri, 13 Dec 2013 22:24:56 +0000 (14:24 -0800)]
gpu: ion: delete ion_system_mapper.c

The mapper abstraction layer was removed before the initial ion
commit, but a stray ion_system_mapper.c file was left in.  Delete
it.

Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: add compat_ioctl
Rom Lemarchand [Fri, 13 Dec 2013 22:24:55 +0000 (14:24 -0800)]
ion: add compat_ioctl

Add a compat_ioctl to the ion driver

Signed-off-by: Rom Lemarchand <romlem@google.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: change ion_user_handle_t definition to int
Rom Lemarchand [Fri, 13 Dec 2013 22:24:54 +0000 (14:24 -0800)]
ion: change ion_user_handle_t definition to int

Turn ion_user_handle_t to int. This change reflects the underlying type
returned by the ion driver.

Signed-off-by: Rom Lemarchand <romlem@google.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: add new ion_user_handle_t type for the user-space token
Rom Lemarchand [Fri, 13 Dec 2013 22:24:53 +0000 (14:24 -0800)]
ion: add new ion_user_handle_t type for the user-space token

Declare new ion_user_handle_t type to contain the token returned to user-space.
This allows a 2-step migration of the user-space code to a new kernel header
first, then will allow us to change the definition of the ion_user_handle_type_t
to int without breaking the API.

Signed-off-by: Rom Lemarchand <romlem@google.com>
(cherry picked from commit ebb8269bbb05b06ecedca3e21b3e65f23d48eadd)
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: don't use id 0 for handle cookie
Colin Cross [Fri, 13 Dec 2013 22:24:52 +0000 (14:24 -0800)]
ion: don't use id 0 for handle cookie

ion userspace clients think that the cookie is a pointer, so they
use NULL to check if the handle has been initialized.  Set the first
id number to 1.

Signed-off-by: Colin Cross <ccross@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: index client->handles rbtree by buffer
Colin Cross [Fri, 13 Dec 2013 22:24:51 +0000 (14:24 -0800)]
ion: index client->handles rbtree by buffer

The only remaining users of the client->handles rbtree are
iterating through it like a list.  Keep the rbtree, but change
its index to be the buffer address instead of the handle address,
which makes ion_handle_lookup a fast rbtree search.

Signed-off-by: Colin Cross <ccross@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: replace userspace handle cookies with idr
Colin Cross [Fri, 13 Dec 2013 22:24:50 +0000 (14:24 -0800)]
ion: replace userspace handle cookies with idr

Userspace handles should not leak kernel virtual addresses to
userspace.  They have to be validated by looking them up in an
rbtree anyways, so replace them with an idr and validate them
by using idr_find to convert the id number to the struct
ion_handle pointer.

Signed-off-by: Colin Cross <ccross@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: remove IS_ERR_OR_NULL
Colin Cross [Fri, 13 Dec 2013 22:24:49 +0000 (14:24 -0800)]
ion: remove IS_ERR_OR_NULL

IS_ERR_OR_NULL is often part of a bad pattern that can accidentally
return 0 on error:
if (IS_ERR_OR_NULL(ptr))
    return PTR_ERR(ptr);

It also usually means that the errors of a function are not well
defined.  Replace all uses in ion.c by ensure that the return
type of any function in ion is an ERR_PTR.

Specify that the expected return value from map_kernel or map_dma
heap ops is ERR_PTR, and warn if a heap returns NULL.

Signed-off-by: Colin Cross <ccross@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: convert map_kernel to return ERR_PTR
Colin Cross [Fri, 13 Dec 2013 22:24:48 +0000 (14:24 -0800)]
ion: convert map_kernel to return ERR_PTR

ion is going to stop accepting NULL as an error value, use ERR_PTR.

Signed-off-by: Colin Cross <ccross@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: add free list size to heap debug files
Colin Cross [Fri, 13 Dec 2013 22:24:47 +0000 (14:24 -0800)]
ion: add free list size to heap debug files

Signed-off-by: Colin Cross <ccross@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: chunk_heap: fix leak in allocated counter
Colin Cross [Fri, 13 Dec 2013 22:24:46 +0000 (14:24 -0800)]
ion: chunk_heap: fix leak in allocated counter

buffer->size is controlled by the outer ion layer, don't modify it
inside the heap.  Instead, compute the rounded up allocated size
on demand.

Signed-off-by: Colin Cross <ccross@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix performance issue in faulting code
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:45 +0000 (14:24 -0800)]
gpu: ion: Fix performance issue in faulting code

Previously the code to fault ion buffers in one page at a time had a
performance problem caused by the requirement to traverse the sg list
looking for the right page to load in (a result of the fact that the items in
the list may not be of uniform size).  To fix the problem, for buffers
that will be faulted in, also keep a flat array of all the pages in the buffer
to use from the fault handler.  To recover some of the additional memory
footprint this creates per buffer, dirty bits used to indicate which
pages have been faulted in to the cpu are now stored in the low bit of each
page struct pointer in the page array.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: add CMA heap
Benjamin Gaignard [Fri, 13 Dec 2013 22:24:44 +0000 (14:24 -0800)]
gpu: ion: add CMA heap

New heap type ION_HEAP_TYPE_DMA where allocation is done with dma_alloc_coherent API.
device coherent_dma_mask must be set to DMA_BIT_MASK(32).
ion_platform_heap private field is used to retrieve the device linked to CMA,
if NULL the default CMA area is used.
ion_cma_get_sgtable is a copy of dma_common_get_sgtable function which should
be in kernel 3.5

Signed-off-by: Benjamin Gaignard <benjamin.gaignard@linaro.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: fix ion_platform_data definition
Benjamin Gaignard [Fri, 13 Dec 2013 22:24:43 +0000 (14:24 -0800)]
gpu: ion: fix ion_platform_data definition

fix ion_platform_heap to make is use an usual way in board configuration file.

Signed-off-by: Benjamin Gaignard <benjamin.gaignard@linaro.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agoion: Add Kconfig dependency to ARM
John Stultz [Fri, 13 Dec 2013 22:24:42 +0000 (14:24 -0800)]
ion: Add Kconfig dependency to ARM

The ion code has some very specific arm-isms which keeps it
from building on other architectures. These should probably be
resolved, but in the mean time, add a dependency on CONFIG_ARM
to avoid build failures.

v2: Fix earlier flub, sending out an early untested version of
the patch.

Cc: Arve Hjønnevåg <arve@android.com>
Cc: Rebecca Schultz Zavin <rebecca@android.com>
Cc: Android Kernel Team <kernel-team@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Remove __GFP_NO_KSWAPD
Arve Hjønnevåg [Fri, 13 Dec 2013 22:24:41 +0000 (14:24 -0800)]
gpu: ion: Remove __GFP_NO_KSWAPD

It no longer exists.

Signed-off-by: Arve Hjønnevåg <arve@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: __dma_page_cpu_to_dev -> arm_dma_ops.sync_single_for_device hack
Arve Hjønnevåg [Fri, 13 Dec 2013 22:24:40 +0000 (14:24 -0800)]
gpu: ion: __dma_page_cpu_to_dev -> arm_dma_ops.sync_single_for_device hack

Signed-off-by: Arve Hjønnevåg <arve@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Also shrink memory cached in the deferred free list
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:39 +0000 (14:24 -0800)]
gpu: ion: Also shrink memory cached in the deferred free list

When the system is low on memory, we want to shrink any cached
system memory ion is holding.  Previously we were shrinking memory
in the page pools, but not in the deferred free list.  This patch
makes it possible to shrink both.  It also moves the shrinker
code into the heaps so they can correctly manage any caches they
might contain.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix bug in ion shrinker
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:38 +0000 (14:24 -0800)]
gpu: ion: Fix bug in ion shrinker

The high variable was sometimes used uninitialized

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: ion_chunk_heap: Zero chunk heap memory at creation time
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:37 +0000 (14:24 -0800)]
gpu: ion: ion_chunk_heap: Zero chunk heap memory at creation time

Allocations from the ion heap need to be zeroed to protect userspace
from seeing memory belonging to other processes.  First allocations
from this heap were not zero'd allowing users to see memory from other
processes on a warm reset.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: fix kfree/list_del order
JP Abgrall [Fri, 13 Dec 2013 22:24:36 +0000 (14:24 -0800)]
gpu: ion: fix kfree/list_del order

With CONFIG_SLUB_DEBUG_ON it would panic during
ion_alloc()
 ion_buffer_create()
   io_heap_drain_freelist()

Signed-off-by: JP Abgrall <jpa@google.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Make ion_free asynchronous
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:35 +0000 (14:24 -0800)]
gpu: ion: Make ion_free asynchronous

Add the ability for a heap to free buffers asynchrounously.  Freed buffers
are placed on a free list and freed from a low priority background thread.
If allocations from a particular heap fail, the free list is drained.  This
patch also enable asynchronous frees from the chunk heap.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add support for sharing buffers with dma buf kernel handles
Johan Mossberg [Fri, 13 Dec 2013 22:24:34 +0000 (14:24 -0800)]
gpu: ion: Add support for sharing buffers with dma buf kernel handles

Currently ion can only share buffers with dma buf fd's. Fd's can not be
used inside the kernel as they are process specific so support for
sharing buffers with dma buf kernel handles is needed to support kernel
only use cases. An example use case could be a GPU driver using ion
that wants to share its output buffers with a 3d party display
controller driver supporting dma buf.

Signed-off-by: Johan Mossberg <johan.mossberg@stericsson.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Only flush buffers in the chunk heap if they were used cached
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:33 +0000 (14:24 -0800)]
gpu: ion: Only flush buffers in the chunk heap if they were used cached

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Refactor the code to zero buffers
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:32 +0000 (14:24 -0800)]
gpu: ion: Refactor the code to zero buffers

Refactor the code in the system heap used to map and zero the buffers
into a seperate utility so it can be called from other heaps.  Use it from
the chunk heap.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Modify zeroing code so it only allocates address space once
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:31 +0000 (14:24 -0800)]
gpu: ion: Modify zeroing code so it only allocates address space once

vmap/vunmap spend a significant amount of time allocating the
address space to map into.  Rather than allocating address space
for each page, instead allocate once for the entire allocation
and then just map and unmap each page into that address space.

Signed-off-by: Rebecca Schultz Zavin <rschultz@google.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Remove heapmask from client
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:30 +0000 (14:24 -0800)]
gpu: ion: Remove heapmask from client

The heapmask in the client generally wasn't being used.  This
patch removes it.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Export ion_client_create
Johan Mossberg [Fri, 13 Dec 2013 22:24:29 +0000 (14:24 -0800)]
gpu: ion: Export ion_client_create

Will enable modules to allocate memory with ion.

Signed-off-by: Johan Mossberg <johan.mossberg@stericsson.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Clarify variable names and comments around heap ids v types
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:28 +0000 (14:24 -0800)]
gpu: ion: Clarify variable names and comments around heap ids v types

There is some confusion between when to use the heap type and when
the id.  This patch clarifies this by using clearer variable names
and describing the intention in the comments.  Also fixes the client
debug code to print heaps by id instead of type.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add chunk heap
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:27 +0000 (14:24 -0800)]
gpu: ion: Add chunk heap

This patch adds support for a chunk heap that allows for buffers that are
made up of a list of fixed size chunks taken from a carveout.  Chunk sizes
are configured when the heaps are created by passing the chunk size in the
priv field of the heap platform data.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Refactor common mapping functions out of system heap
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:26 +0000 (14:24 -0800)]
gpu: ion: Refactor common mapping functions out of system heap

The system heap contained several general purpose functions to map
buffers to the kernel and userspace.  This patch refactors those
into ion_heap.c so they can be used by other heaps.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Switch heap rbtree to a prio list
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:25 +0000 (14:24 -0800)]
gpu: ion: Switch heap rbtree to a prio list

Switches the rbtree tree of heaps for a plist.  This significantly
simplifies the code and the list is small and is modified only at
first boot so the rbtree is unnecessary.  This also switches
the traversal of the heap list to traverse from highest to lowest
id's.  This allows allocations to pass a heap mask that falls
back on the system heap -- typically id 0, which is the common case.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix bug where MAP ioctl was no longer supported
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:24 +0000 (14:24 -0800)]
gpu: ion: Fix bug where MAP ioctl was no longer supported

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Modify reserve function for carveouts with no start address
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:23 +0000 (14:24 -0800)]
gpu: ion: Modify reserve function for carveouts with no start address

This patch allows you to specify a heap that requires carveout memory
but that doesn't specify a start address.  Memblock_alloc will be called
to find a location for these heaps.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: fix compilation warning
Benjamin Gaignard [Fri, 13 Dec 2013 22:24:22 +0000 (14:24 -0800)]
gpu: ion: fix compilation warning

use atomic_read to get the refcount value to avoid compilation warning

Signed-off-by: Benjamin Gaignard <benjamin.gaignard@linaro.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: fix carveout ops
Benjamin Gaignard [Fri, 13 Dec 2013 22:24:21 +0000 (14:24 -0800)]
gpu: ion: fix carveout ops

when using carveout heap ion_buffer_create function failed because
map_dma and unmap_dma operations aren't set by carveout heap.

Signed-off-by: Benjamin Gaignard <benjamin.gaignard@linaro.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix bug in zeroing pages in system heap
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:20 +0000 (14:24 -0800)]
gpu: ion: Fix bug in zeroing pages in system heap

Pages are zeroed for security purposes when returned to the
ion heap.  There was a bug in this code preventing this
from happening.

Bug: 7573871
Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix bug in ion_system_heap map_user
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:19 +0000 (14:24 -0800)]
gpu: ion: Fix bug in ion_system_heap map_user

When the requested mmap length was not an integer number of
chunks or the buffer, or if an offset was provided, a bug
would cause extra or incorrect pages of the buffer to be mapped.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Don't flush allocatoins that come from the page pools
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:18 +0000 (14:24 -0800)]
gpu: ion: Don't flush allocatoins that come from the page pools

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Clear GFP_WAIT flag on high order allocations
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:17 +0000 (14:24 -0800)]
gpu: ion: Clear GFP_WAIT flag on high order allocations

This will prevent the kernel from kicking off compaction
when higher order allocations are made.  Instead we will
get these high order allocations only if they are readily
available.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Refactor locking
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:16 +0000 (14:24 -0800)]
gpu: ion: Refactor locking

Removes contention for lock between allocate and free by reducing
the length of time the lock is held for.  Split out a seperate
lock to protect the list of heaps and replace it with a rwsem since
the list will most likely only be updated during initialization.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Switch to using a single shrink function
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:15 +0000 (14:24 -0800)]
gpu: ion: Switch to using a single shrink function

The single shrink function will free lower order pages first. This
enables compaction to work properly.

[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix lockdep issue in ion_page_pool
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:14 +0000 (14:24 -0800)]
gpu: ion: Fix lockdep issue in ion_page_pool

Currently the mutex is held while kmalloc is called, under a low memory
condition this might trigger the shrinker which also takes this mutex.
Refactor so the mutex is not held during allocation.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix several issues with page pool
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:13 +0000 (14:24 -0800)]
gpu: ion: Fix several issues with page pool

Split out low and high mem pages so they are correctly reported
when the shrinker is called.
Fix potential deadlock caused by holding the page pool lock while
allocationg and also needing that lock from the shrink function

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Modify gfp flags in ion_system_heap
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:12 +0000 (14:24 -0800)]
gpu: ion: Modify gfp flags in ion_system_heap

When allocations larger than order 4 are made, use _GFP_NORETRY
and __GFP_NO_KSWAPD so kswapd doesn't get kicked off to reclaim
these larger chunks.  For smaller allocaitons, these are
unnecessary, as the system should be able to reclaim these.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Use the ion_page_pool from the system heap
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:11 +0000 (14:24 -0800)]
gpu: ion: Use the ion_page_pool from the system heap

With this change the system heap will use pagepools to avoid
having to invalidate memory when it is allocated, a
significant performance improvement on some systems.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add ion_page_pool.
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:10 +0000 (14:24 -0800)]
gpu: ion: Add ion_page_pool.

This patch adds a new utility heaps can use to manage
memory.  In the past we have found it can be very
expensive to manage the caches when allocating memory,
but it is imposible to know whether a previous user of a
given memory allocation had a cached mapping. This patch
adds the ability to store a pool of pages that were
previously used uncached so that cache maintenance
only need be done when growing this pool.  The pool also
contains a shrinker so memory from the pool can be
recovered in low memory conditions.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: use vmalloc to allocate page array to map kernel
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:09 +0000 (14:24 -0800)]
gpu: ion: use vmalloc to allocate page array to map kernel

When ion_map_kernel is execute the system must allocate
an array large enough to hold a pointer to each page in
the buffer.  If the buffer is very large and the system
memory has become very fragmented, there may not be
sufficient high order allocations available from kmalloc.
Use vmalloc instead.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: ion_system_heap: Fix bug preventing compilation
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:08 +0000 (14:24 -0800)]
gpu: ion: ion_system_heap: Fix bug preventing compilation

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Stop trying to allocate from an order on first failure
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:07 +0000 (14:24 -0800)]
gpu: ion: Stop trying to allocate from an order on first failure

With this patch the system heap will only try to allocate from each
order as long as allocations succeed.  If it failes to obtain a higher
order allocation, it doesn't retry that order.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: optimize system heap for non fault buffers
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:06 +0000 (14:24 -0800)]
gpu: ion: optimize system heap for non fault buffers

If a buffer's user mappings are not going to be faulted
in it need not be allocated page wise.  We can optimize
this common case by allocating an sglist of larger chunks
rather than creating an entry for each page in the
allocation.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add support for cached mappings that don't fault
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:05 +0000 (14:24 -0800)]
gpu: ion: Add support for cached mappings that don't fault

We have found that faulting in the mappings for cached
allocations has a significant performance impact and is
only a benefit if only a small part of the buffer is
touched by the cpu (an uncommon case for software rendering).
This patch introduces a ION_FLAG_CACHED_NEEDS_SYNC
which determines whether a mapping should be created by
faulting or at mmap time.  If this flag is set,
userspace must manage the caches explictly using the SYNC ioctl.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix memory leak of dirty bits
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:04 +0000 (14:24 -0800)]
gpu: ion: Fix memory leak of dirty bits

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add debug information for orphaned handles
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:03 +0000 (14:24 -0800)]
gpu: ion: Add debug information for orphaned handles

It is possible for a buffer to exist only as a dma_buf file
descriptor without it being held in any handles.  When this
occurs it is impossible to track where the buffer is in the
system (without traversing every process in the system and
inspecting its file table).  When buffers are orphaned like
this, copy the task comm and pid of the last client to hold
them into the buffer so we have a debugging hint as to where
this buffer came from.  In practice this will probalby be
the process that allocated the buffer.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix bug in ion_free
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:02 +0000 (14:24 -0800)]
gpu: ion: Fix bug in ion_free

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Fix race between ion_import and ion_free
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:24:01 +0000 (14:24 -0800)]
gpu: ion: Fix race between ion_import and ion_free

If preemted during ion_free after the refcount is updated but
before the handle can be removed from the rb_tree, import
might find that handle in the tree and try to reuse it
when execution returns to free, the handle will be cleaned
up leaving the caller of import with a corrupt handle.
This patch modifies the locking to protect agains this race.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: fix page offset in dma_buf_kmap()
Greg Hackmann [Fri, 13 Dec 2013 22:24:00 +0000 (14:24 -0800)]
gpu: ion: fix page offset in dma_buf_kmap()

Signed-off-by: Greg Hackmann <ghackmann@google.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Switch to using kmalloc rather than kmap during allocation
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:59 +0000 (14:23 -0800)]
gpu: ion: Switch to using kmalloc rather than kmap during allocation

Previously, metadata was stored in the allocated pages themselves
during allocation.  However the system can only have a limited
number of kmapped pages.  A very large allocation might exceed
this limit.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Only map as much of the vma as the user requested
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:58 +0000 (14:23 -0800)]
gpu: ion: Only map as much of the vma as the user requested

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Don't call ion_buffer_put on error path
Laura Abbott [Fri, 13 Dec 2013 22:23:57 +0000 (14:23 -0800)]
gpu: ion: Don't call ion_buffer_put on error path

If dma_buf_fd fails, the dma_buf needs to be cleaned up by
calling dma_buf_put. dma_buf_put will call ion_dma_buf_release
which in turn calls ion_buffer_put to clean up the buffer
reference. Calling ion_buffer_put after dma_buf_put drops the
reference count by one more which is incorrect. Fix this by
getting rid of the extra ion_buffer_put call.

Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: IOCTL return success when error occurs
Olav Haugan [Fri, 13 Dec 2013 22:23:56 +0000 (14:23 -0800)]
gpu: ion: IOCTL return success when error occurs

ION_IOC_MAP, ION_IOC_SHARE, and ION_IOC_IMPORT may return
success when an error occurs.

Add correct error handling to ION_IOC_MAP, ION_IOC_SHARE, and
ION_IOC_IMPORT.

Signed-off-by: Olav Haugan <ohaugan@codeaurora.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add EXPORT_SYMBOL to functions
Olav Haugan [Fri, 13 Dec 2013 22:23:55 +0000 (14:23 -0800)]
gpu: ion: Add EXPORT_SYMBOL to functions

Several functions in the ion interface is missing
EXPORT_SYMBOL. This is needed to allow clients to
use these functions from kernel modules.

Add EXPORT_SYMBOL to functions that are supposed
to be exposed.

Signed-off-by: Olav Haugan <ohaugan@codeaurora.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add missing argument to WARN call
Olav Haugan [Fri, 13 Dec 2013 22:23:54 +0000 (14:23 -0800)]
gpu: ion: Add missing argument to WARN call

The condition argument to the WARN call in ion_free and
ion_share_dma_buf are missing. Add the argument to
allow correct printing of warning message.

Signed-off-by: Olav Haugan <ohaugan@codeaurora.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: do not ask for compound pages in system heap
Dima Zavin [Fri, 13 Dec 2013 22:23:53 +0000 (14:23 -0800)]
gpu: ion: do not ask for compound pages in system heap

Signed-off-by: Dima Zavin <dima@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add explicit sync ioctl
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:52 +0000 (14:23 -0800)]
gpu: ion: Add explicit sync ioctl

This is deprecated in favor of using the dma_buf api which will
automatically sync a buffer to memory when it is mapped to a device.
However, that functionality is not ready, so this patch adds the
ability to sync a buffer explicitly.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Modify the system heap to try to allocate large/huge pages
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:51 +0000 (14:23 -0800)]
gpu: ion: Modify the system heap to try to allocate large/huge pages

On some systems there is a performance benefit to reducing tlb pressure
by minimizing the number of chunks in an allocation.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Add cache maintenance to ion.
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:50 +0000 (14:23 -0800)]
gpu: ion: Add cache maintenance to ion.

This patch adds cache maintenance operations to ion.  As per mailing
list discussions regarding dma_buf, cache operations are done implicitly.
At buffer allocaiton time the user can select whether he'd like mappings
(both kernel and user) to be cached.  When cached mappings are selected,
no mappings will be created for a buffer at mmap time.  Instead pages will
be faulted in one at a time so we can track which pages require flushing
before dma.  When the buffers are mapped for dma (via the dma_buf apis)
any pages which were touched will be synced for device.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Map only the vma size given
Laura Abbott [Fri, 13 Dec 2013 22:23:49 +0000 (14:23 -0800)]
gpu: ion: Map only the vma size given

When mapping carveout buffers into userspace, only map
the size of the vma given, not the full size of the buffer
since clients may map less than the buffer size.

Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Loop on the handle count when destroying
Laura Abbott [Fri, 13 Dec 2013 22:23:48 +0000 (14:23 -0800)]
gpu: ion: Loop on the handle count when destroying

When destroying a handle, all kernel mappings to that handle
should be destroyed. Other handles may still have references
and valid mappings to the buffer underneath which should not
be destroyed. Loop on the handle reference count, not the buffer
reference count to get rid of all kernel mappings for the handle.

Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: ion_system_heap: Change allocations to GFP_HIGHUSER
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:47 +0000 (14:23 -0800)]
gpu: ion: ion_system_heap: Change allocations to GFP_HIGHUSER

Previously the ion_system_heap was using GFP_KERNEL, forcing all allocations
to be in lowmem.  This quickly causes us to run out of lowmem.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Set the dma_address of the sg list at alloc time
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:46 +0000 (14:23 -0800)]
gpu: ion: Set the dma_address of the sg list at alloc time

This patch sets the dma_address field of the sglist representing
an allocation at allocation time.  This technically breaks the dma api
which states that these addresses should be set when a particular device
takes ownership of a buffer via the dma_map apis.  In the case of our
systems the only dma address space is physical addresses.  Additionally,
we can not afford the overhead of calling dma_map_sg from this location
as it implies a cache invalidate that is not necessary if the memory
was previously mapped cached.  Instead, the expectation is that memory
being returned from the heaps is ready for dma in that if any cached
mappings of that memory exist they have been invalidated.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: fill in buffer->{dev,size} before mapping new buffers
Greg Hackmann [Fri, 13 Dec 2013 22:23:45 +0000 (14:23 -0800)]
gpu: ion: fill in buffer->{dev,size} before mapping new buffers

At least one map_dma() implementation (EXYNOS_CONTIG) assumes the fields
are filled in

Signed-off-by: Greg Hackmann <ghackmann@google.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Get an sg_table from an ion handle
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:44 +0000 (14:23 -0800)]
gpu: ion: Get an sg_table from an ion handle

This patch adds an interface to return and sg_table given a
valid ion handle.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: Allocate the sg_table at creation time rather than dynamically
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:43 +0000 (14:23 -0800)]
gpu: ion: Allocate the sg_table at creation time rather than dynamically

Rather than calling map_dma on the allocations dynamically, this patch
switches to creating the sg_table at the time the buffer is created.
This is necessary because in future updates the sg_table will be used
for cache maintenance.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
10 years agogpu: ion: support begin/end and kmap/kunmap dma_buf ops
Rebecca Schultz Zavin [Fri, 13 Dec 2013 22:23:42 +0000 (14:23 -0800)]
gpu: ion: support begin/end and kmap/kunmap dma_buf ops

These ops were added in the 3.4 kernel.  This patch adds support
for them to ion.  Previous ion_map/unmap_kernel api is also
retained in addition to this new api.

Signed-off-by: Rebecca Schultz Zavin <rebecca@android.com>
[jstultz: modified patch to apply to staging directory]
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>