Re: [PATCH v9 0/13] implement dlmem() function

[Szabolcs Nagy <szabolcs.nagy@arm.com>]

The 03/29/2023 18:18, stsp via Libc-alpha wrote:
> Just to be more constructive, here is the
> example:
> 
> void *dlopen_with_offset(const char *file, off_t offset, int flags)
> {

this api requires no libc change:

dump the contents to a temp file and use dlopen on it.


>     off_t len;
>     void *addr;
>     void *handle;
>     int fd;
> 
>     fd = open(file, O_RDONLY);
>     if (fd == -1)
>         return NULL;
>     len = lseek(fd, 0, SEEK_END);
>     lseek(fd, 0, SEEK_SET);
>     if (len <= offset)
>       goto err_close;
>     len -= offset;
>     /* if offset unaligned then just use read() */
>     if (offset & (PAGE_SIZE - 1)) {
>         int rd;
>         addr = mmap(NULL, len, PROT_READ | PROT_WRITE,
>                     MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
>         if (addr == MAP_FAILED)
>             goto err_close;
>         rd = read(fd, addr, len);
>         if (rd != len) {
>             munmap(addr, len);
>             goto err_close;
>         }
>     } else {
>         addr = mmap(NULL, len, PROT_READ, MAP_PRIVATE, fd, offset);
>     }
>     close(fd);
>     if (addr == MAP_FAILED)
>         return NULL;
>     handle = dlmem(addr, len, flags, NULL);
>     munmap(addr, len);
>     return handle;
> 
> err_close:
>     close(fd);
>     return NULL;
> }
> 
> 
> Does it parse an elf headers?
> If so - where?

it creates a mapping that's not compatible with dynamic loading
requirements *depending on* what the elf headers say.
(elf segments may be loaded at various offsets that can conflict
with the mapping that you just made)

but i'm not going to think about all the different ways this
is broken: you cut across abstraction layers exposing dynamic
loading internals where a temp file works fine. if there is
a performance problem (e.g. page sharing) likely that can be
solved too via kernel level apis, not ld.so.