23668a4b9b
Signed-off-by: Xiang Xiao <xiaoxiang@xiaomi.com> |
||
---|---|---|
.. | ||
fs_mmap.c | ||
fs_munmap.c | ||
fs_rammap.c | ||
fs_rammap.h | ||
Kconfig | ||
Make.defs | ||
README.txt |
fs/mmap README File =================== NuttX operates in a flat open address space and is focused on MCUs that do support Memory Management Units (MMUs). Therefore, NuttX generally does not require mmap() functionality and the MCUs generally cannot support true memory-mapped files. However, memory mapping of files is the mechanism used by NXFLAT, the NuttX tiny binary format, to get files into memory in order to execute them. mmap() support is therefore required to support NXFLAT. There are two conditions where mmap() can be supported: 1. mmap can be used to support eXecute In Place (XIP) on random access media under the following very restrictive conditions: a. The filesystem supports the FIOC_MMAP ioctl command. Any file system that maps files contiguously on the media should support this ioctl. (vs. file system that scatter files over the media in non-contiguous sectors). As of this writing, ROMFS is the only file system that meets this requirement. b. The underlying block driver supports the BIOC_XIPBASE ioctl command that maps the underlying media to a randomly accessible address. At present, only the RAM/ROM disk driver does this. Some limitations of this approach are as follows: a. Since no real mapping occurs, all of the file contents are "mapped" into memory. b. All mapped files are read-only. c. There are no access privileges. 2. If CONFIG_FS_RAMMAP is defined in the configuration, then mmap() will support simulation of memory mapped files by copying files whole into RAM. These copied files have some of the properties of standard memory mapped files. There are many, many exceptions, however. Some of these include: a. The goal is to have a single region of memory that represents a single file and can be shared by many threads. That is, given a filename a thread should be able to open the file, get a file descriptor, and call mmap() to get a memory region. Different file descriptors opened with the same file path should get the same memory region when mapped. The limitation in the current design is that there is insufficient knowledge to know that these different file descriptors correspond to the same file. So, for the time being, a new memory region is created each time that rammap() is called. Not very useful! b. The entire mapped portion of the file must be present in memory. Since it is assumed that the MCU does not have an MMU, on-demanding paging in of file blocks cannot be supported. Since the while mapped portion of the file must be present in memory, there are limitations in the size of files that may be memory mapped (especially on MCUs with no significant RAM resources). c. All mapped files are read-only. You can write to the in-memory image, but the file contents will not change. d. There are no access privileges. e. Since there are no processes in NuttX, all mmap() and munmap() operations have immediate, global effects. Under Linux, for example, munmap() would eliminate only the mapping with a process; the mappings to the same file in other processes would not be effected. f. Like true mapped file, the region will persist after closing the file descriptor. However, at present, these ram copied file regions are *not* automatically "unmapped" (i.e., freed) when a thread is terminated. This is primarily because it is not possible to know how many users of the mapped region there are and, therefore, when would be the appropriate time to free the region (other than when munmap is called). NOTE: Note, if the design limitation of a) were solved, then it would be easy to solve exception d) as well.