- Arguments and flags
- Sharing memory with MMAP
- MISC: MMAP is faster than reading a file in blocks
- Future readings
MMAP is a UNIX system call that maps files into memory. It’s a method used for memory-mapped file I/O. It brings in the optimization of
lazy loading or
demand paging such that the I/O or reading file doesn’t happen when the memory allocation is done, but when the memory is accessed. After the memory is no longer needed it can be cleared with
MMAP supports certain flags or argument which makes it suitable for allocating memory without file mapping as well. In Linux kernel, the
malloc call uses
MAP_ANONYMOUS flag for large allocations.
In this article, I’ll be explaining how what
mmap is and how it can be used for sharing memory in Linux. It kind of is the backbone of shared memory in Android.
Arguments and flags
mmap() creates a new mapping in the virtual address space of the calling process. If you check out the Linux kernel page for mmap you’ll see several arguments and flags. On the other hand,
munmap() is used to free the allocated memory.
addrspecifies the starting address of the allocation and if it’s passed as
NULLthe kernel chooses the starting address.
lengthargument specifies the length of allocation in bytes and should be
protargument describes the protection level
PROT_EXECPages may be executed.
PROT_READPages may be read.
PROT_WRITEPages may be written.
PROT_NONEPages may be not be accessed.
The flags can be passed with bitwise OR operator and the default protection level is
PROT_EXEC | PROT_READ | PROT_WRITE
flagsargument is used to do various customizations, the interesting ones are:
MAP_PRIVATEdetermines if the pages would be shared across processes or owned by a single process.
MAP_ANON) is used to indicate that the pages are not backed by any file and are anonymous.
malloccould use this flag to create a large memory allocation.
fdindicates the file descriptor that represents the file backing the memory. This is not needed when using
MAP_ANONflag and can be set to
offsetindicates the offset in bytes from which the file should be read to memory. The allocated memory will load
offset + lengthbytes from the file when the memory access happens.
munmap()system call deletes the mappings for the specified address range and causes further references to addresses within the range to generate invalid memory references. The region is also automatically unmapped when the process is terminated. On the other hand, closing the file descriptor does not unmap the region.
int munmap(void *addr, size_t length);
addris the address of allocation to free, essentially what you got from calling the
mmap(). After calling
munmap(), any access on the memory address shall raise
lengthdetermines the area of memory to clear. The area of memory from
addr + lengthwould be freed on this call.
Sharing memory with MMAP
MMAP can be thought of as the core memory allocation API in Linux and several high-level constructs take advantage of this for providing various features. Linux kernel is the core of Android OS and components like
MMAP in its core.
ASHMEM is used for sharing memory in Android in different components like ContentProviders or Binder IPC.
Sharing between parent and child
This is fairly simple to visualize. A
mmap allocation with
MAP_SHARED flag can be accessed directly by the child process.
This is very helpful in sharing the memory of core components in Android. All applications in Android are forked from a bare-bone process called
Zygote which loads the core libraries and code required by all applications with
Zygote is loaded into memory on device boot and when a user attempts to open an application for the first time the system forks
Zygote and then the application logic is initialized.
Sharing between siblings
While it’s easy to visualize how memory can be shared in ancestry between a parent and child. The logic is very similar but involves Inter-Process Communication (IPC). Two common ways to achieve this could be:
Without extra management layer
The concept is similar, the two processes say
Process 1 and
Process 2 can communicate with each other via certain IPC technology.
Process 1creates a file and allocates memory on that with
MAP_SHAREDflag and appropriate protection level and length. This process can write some data in the allocated memory space.
Process 1shares this file descriptor with Process 2 via a certain IPC method.
Process 2receives this file descriptor and calls
mmapon that. So the system returns the virtual address of the same memory allocation and based on the protection levels set by
Process 2can read, write or execute the shared memory pages.
However, these processes are responsible for explicitly deallocating memory, otherwise, it cannot be reused by another process in need of memory.
With an extra management layer
In this case, another process acts as the manager of shared memory and exposes interface or methods to allocate or retrieve memory allocations. Let’s say there is a memory manager called
XMAN and exposes APIs like this:
Process 1could allocate a chunk of memory using
Xman_allocate()and share the
Xman_allocation.fdwith another process via a certain IPC mechanism.
Process 2could use
Xman_get()to get the same allocation and act on it.
- Any of these processes could use the
Xman_free()to explicitly free the memory.
While the way of dealing with shared memory seems very similar with or without a manager instance, a centralized manager can abstract some memory freeing techniques thus taking away the expectation of being good citizens from the calling processes like:
- Freeing memory after use, the Manager can take care of freeing when the calling processes die.
- Some components like ASHMEM, support features like PINNING and UNPINNING section of memory which allows the consumer process to set which part of memory can be cleared when the system is out of free memory. This protects the consumer apps from being killed by the Low Memory Killer (LMK) when it’s reclaiming memory. ASHMEM has its process on deciding which UNPINNED memory to clear when available memory is system goes below a certain threshold.
MISC: MMAP is faster than reading a file in blocks
While exploring these concepts I was wondering how file-backed memory manages to be performant while file IO operation like
read() is generally considered much slower than memory operations. There are a few interesting StackOverflow questions like Why mmap() is faster than sequential IO? and mmap() vs. reading blocks which answer this questions pretty well.
But they won’t give you a short answer like - Because MMAP is magic! They are long reads.
I wish I could add a TL;DR; answer to this question here but there isn’t one. Both
read() have their pros and cons and could be more performant in different situations. While
mmap() seems like magic, it’s simply not.
In the future, I intend to write about what ASHMEM is, how it works, why it was brought when MMAP existed and examples of how it’s been used in Android. Another interesting memory manager in Android is the ION memory manager which was added to Linux kernel in 2011 by a patch from Google to solve issues around large memory allocations needed by components like GPU, display, camera, etc.
Want to read more such similar contents?
I like to write articles on topic less covered on internet. They revolve around writing fast algorithms, image processing as well as general software engineering.
I publish many of them on Medium.
If you are already on medium - Please join 4200+ other members and Subscribe to my articles to get updates as I publish.
If you are not on Medium - Medium has millions of amazing articles from 100K+ authors. To get access to those, please join using my referral link. This will give you access to all the benefits of Medium and Medium shall pay me a piece to support my writing!