As has been observed in #28 (comment), #23, #8, and conversations elsewhere, wasi-filesystems rights system is complex to implement, complex to use, and not very valuable in practice. It doesn't reflect anything in POSIX, so it doesn't give us any extra POSIX compatibility. It requires extra code in libc, both in path_open to translate POSIX open flags into rights flags, and to translate from ENOTCAPABLE into POSIX errno values, and we care about code size!
fdstat_t and path_open will need a new value to specifying the file mode, which can be either read, write, or read+write.
fd_fdstat_set_rights will no longer be needed. Note that POSIX doesn't have a way to change the read or write flags on an open file descriptor after it's opened, so this won't remove any POSIX functionality.
Note that, as in POSIX, this would mean that directories must be opened with read (and not write) permissions, and the resulting handle can mutate the filesystem (assuming the filesystem ACLs permit it). As in POSIX, this would mean there's no built in concept of a readonly view of an otherwise writeable filesystem. Such a thing would seem useful, however it's not in POSIX, and it is something that will eventually be implementable with API virtualization, so I'm proposing we just let this be unsupported for now.
As has been observed in #28 (comment), #23, #8, and conversations elsewhere, wasi-filesystems rights system is complex to implement, complex to use, and not very valuable in practice. It doesn't reflect anything in POSIX, so it doesn't give us any extra POSIX compatibility. It requires extra code in libc, both in
path_opento translate POSIX open flags into rights flags, and to translate fromENOTCAPABLEinto POSIX errno values, and we care about code size!fdstat_tandpath_openwill need a new value to specifying the file mode, which can be either read, write, or read+write.fd_fdstat_set_rightswill no longer be needed. Note that POSIX doesn't have a way to change the read or write flags on an open file descriptor after it's opened, so this won't remove any POSIX functionality.Note that, as in POSIX, this would mean that directories must be opened with read (and not write) permissions, and the resulting handle can mutate the filesystem (assuming the filesystem ACLs permit it). As in POSIX, this would mean there's no built in concept of a readonly view of an otherwise writeable filesystem. Such a thing would seem useful, however it's not in POSIX, and it is something that will eventually be implementable with API virtualization, so I'm proposing we just let this be unsupported for now.