private_handle_t's memory layout must be consistent between the 32-bit
and 64-bit gralloc, in case buffers are passed between processes.
Replace the (variably sized) uintptr_t base with a fixed size uint64_t,
and enforce 8-byte alignment for architectures where uint64_t alignment
varies between 32-bit and 64-bit.
Change-Id: I06cb31d4b9620ea18e5b50d3a3142b5adb2d2a14
Signed-off-by: Greg Hackmann <ghackmann@google.com>
A constant sNumInts with a value of 6 was being assigned to the numInts
attribute in the private_handle_t constructor, and was also used in the
validate method. That constant value is appropriate for 32-bit systems
but is not appropriate for a 64-bit system where uintptr_t used for
base attribute will be 64-bit.
sNumInts is now changed to a static inline function that calculates
numInts.
Change-Id: I482ddb5915c9ff55fb2e2a87887a0ec2dc2299ed
Signed-off-by: Marcus Oakland <marcus.oakland@arm.com>
Signed-off-by: Ashok Bhat <ashok.bhat@arm.com>
The register/unregister gralloc calls were avoiding
mmapping/munmapping the shared memory region if the buffer was created
by the current process. This is left over from the pmem-based
implementation, where trying to map the same region twice in the same
process would fail, or would reuse a single mapping without
refcounting.
This causes problems if a buffer is
- allocated in process A,
- transferred from A to process B and registered there
- unregistered/freed in A
- transferred back from B to A and re-registered
Process A then has a new handle to the buffer, but since it originally
created the buffer it will not be mmapped, so trying to read or write
the buffer will crash.
With ashmem, mmaping a region twice in the same process creates two
distinct mappings which can be used and munmapped independently. So
we no longer need to avoid mmapping again in the allocating process.
Bug: 8468756
Change-Id: I167bec5ca07e5534c5e2115630fe8386e481388e
this gralloc module is only used on the emulator or without a h/w
renderer. therefore there is no synchronization to do in lock/unlock
and pmem buffers are not relevant.
hopefully this will remove some of the confusion about how gralloc
should be implemented and make it more obvious that this implementation
is not intended to be used by h/w renderers.
The place where the PAGE_SIZE value is defined varies, mostly because
you're not supposed to be using it directly. sysconf(_SC_PAGE_SIZE)
is the approved method, and in fact some Linux distros actually #define
PAGE_SIZE to the library call.
- make sure to return an error if a buffer is locked twice or unlocked while not locked.
- added registerBuffer() and unregisterBuffer() to the gralloc module so that we can do some cleanup when a buffer is no longer needed. this became necessary after we removed map/unmap so we have a place to unmap buffers without the need of a kernel module.
- change the constants for GRALLOC_USAGE_SW_{READ|WRITE}_NEVER to 0, so that NOT specifying them means "NEVER".
