v1

Exploiting race conditions on [ancient] Linux

Jann Horn, Google Project Zero

(if this text is too small for you to read, maybe open the slides on your laptop)

slides at: https://sched.co/TynD

Confidential + Proprietary
Introduction
● bugs described here have been fixed for a long time
● all exploits against kernel 4.4
● focus on exploitation techniques, not impact of individual
bugs

Confidential + Proprietary
Agenda
● physical-page use-after-free via stale TLB [bug 1]
○ [kernel bug, PoC for Google Pixel 2]
○ buddy allocator
○ preemption and scheduler control
● refcount decrement on struct file [bug 2]
○ [kernel bug, PoC for Ubuntu 16.04]
○ userfaultfd() and FUSE
○ kcmp()
● a poor substitute for FUSE/userfaultfd [bug 3]
○ [userspace bug, PoC for Google Pixel 2]
○ i_mutex on kernel 4.4
○ priority inversion
○ repeated file mapping faults
Confidential + Proprietary
Bug 1: mremap()+fallocate() race

Confidential + Proprietary
Translation Lookaside Buffer (TLB)

● in-CPU cache for page table entries (PTEs)

● PTEs are essentially refcounted page pointers
● TLB borrows references from PTEs
● kernel can invalidate TLB for virtual address
ranges
○ x86: IPI for remote CPUs
○ arm64: magic system-wide TLBI instruction
Confidential + Proprietary
mremap(): moving a memory mapping

● moves associated page 1. move PTEs (also

moves references)
table entries (PTEs)
old new
● has to flush the TLB for mapping mapping
the old address range
2. flush TLB
(old mapping
becomes
inaccessible)

Confidential + Proprietary
fallocate(): (de)allocate space for a file

● interesting case: punch a

hole in a file
● file pages in the hole are: mapping

○ yanked out of all mappings

1. delete PTEs
(in all processes) 2. flush TLB
○ released once all references 3. drop references

are gone

Confidential + Proprietary
Bug 1: mremap()+fallocate() race
crbug.com/project-zero/1695

● mremap() holds no relevant lock 1. move PTEs (also

moves references)
between moving PTEs and TLB
flush, fallocate() possible in old new
between mapping mapping

● fallocate() drops page references 5. flush TLB 2. delete PTEs

(old mapping 3. flush TLB
after its TLB flush 4. drop
becomes
● stale TLB entry for old mapping inaccessible) references
(can free pages)
permits physical-page
use-after-free between dropping
page reference and flushing TLB
for old mapping Confidential + Proprietary
Exploit plan: Basics
● biggest impact on Linux <4.9; exploiting
for write access is much harder on
newer kernels
old new
● goal: Pixel 2 (Linux 4.4) exploit mapping mapping
● exploit idea: reallocate freed page with
kernel data TLB flush pending
pages freed
=> physical UAF!

Confidential + Proprietary
 Buddy allocator
percpu freelist [with UAF page] (highly
cpu X simplified, not
MIGRATE_MOVABLE entirely correct)

Page freelist Page freelist Page freelist

order 0 order 1 order 2
MIGRATE_MOVABLE MIGRATE_MOVABLE MIGRATE_MOVABLE

Page freelist Page freelist Page freelist

order 0 order 1 order 2
MIGRATE_UNMOVABLE MIGRATE_UNMOVABLE MIGRATE_UNMOVABLE

percpu freelist SLAB

cpu X e.g. kmalloc-256
MIGRATE_UNMOVABLE Confidential + Proprietary
Exploit plan
● biggest impact on Linux <4.9; exploiting
for write access is much harder on
newer kernels
old new
● goal: Pixel 2 (Linux 4.4) exploit mapping mapping
● exploit idea: reallocate freed page with
kernel data - ✘, looked too messy TLB flush pending
pages freed
● exploit idea: reallocate freed page as => physical UAF!
page cache for privileged code ✔
○ requires disk I/O within the race window
○ need to make the mremap() race window
wide enough for disk I/O
● race window detectable through procfs
Confidential + Proprietary
 CPU 0 CPU 1
Preemption task A task B
wakeup
IPI task C
running blocked running

● waking up a task can cause a scheduler Inter-Processor Interrupt (IPI)

○ depending on policy, priority and past CPU usage [see check_preempt_wakeup() ]
● Linux supports three kernel preemption models:
○ "voluntary" preemption can yield the CPU at cond_resched() [called
in ~1000 places]
■ used by many Linux distributions by default
○ full preemption
■ enabled on Android
■ syscall context interruptible directly via inter-processor interrupt (IPI) on task wakeup
■ no preemption in some code regions (holding a spinlock [/ preemption explicitly disabled
/ interrupts disabled])
■ [preemption requests in critical region are delivered on critical section exit]
■ mutexes don't block preemption!
Confidential + Proprietary
 CPU 0 CPU 1
Scheduler control task A
running
task B
blocked
wakeup
IPI
task C
SCHED_IDLE SCHED_NORMAL running

● sched_setscheduler(): set SCHED_NORMAL / SCHED_IDLE

○ [realtime policies require CAP_SYS_NICE or RealtimeKit]
● on busy CPU, SCHED_IDLE has infrequent wakeups
● SCHED_IDLE never preempts
● sched_setaffinity(): pin task to CPU bitmask
● also affects execution in kernel mode!
➢ pin two own tasks to a single CPU
➢ set different scheduling classes
➢ interrupt kernel code execution

Confidential + Proprietary
 Android kernel exploit (app -> zygote)
task A task B task C task D task E
[pinned to CPU 0] [pinned to CPU 1] [pinned to CPU 1] [pinned to CPU 2] [pinned to CPU 3]
[normal priority] [idle priority] [normal priority] [normal priority] [normal priority]
busyloop
read(<pipe>) busyloop busyloop
y wa it ing
yield b reading procfs attempting to
stats write through
mremap(...) detect old mapping
PTE is moved wakeup mremap() [keep
write(<pipe>)
yield by preemption TLB
entry
sched_setaffinity so that we can block alive]
(task B, CPU 0) without yielding to B
keep
[preempted]
task B fallocate(...) free the pages
off the
CPU pread(<zygote reuse one of
code>) the pages as
page cache overwrite
the page
Confidential + Proprietary
 Android kernel exploit (app -> zygote)
task A task B task C task D task E
[pinned to CPU 0] [pinned to CPU 1] [pinned to CPU 1] [pinned to CPU 2] [pinned to CPU 3]
[normal priority] [idle priority] [normal priority] [normal priority] [normal priority]
busyloop
read(<pipe>) busyloop busyloop
y wa it ing
yield b reading procfs attempting to
stats write through
mremap(...) detect old mapping
PTE is moved wakeup mremap() [keep
write(<pipe>)
yield by preemption TLB
entry
sched_setaffinity so that we can block alive]
(task B, CPU 0) without yielding to B
keep
[preempted]
task B fallocate(...) free the pages
off the
CPU pread(<zygote reuse one of
code>) the pages as
page cache overwrite
the page
Confidential + Proprietary
 Android kernel exploit (app -> zygote)
task A task B task C task D task E
[pinned to CPU 0] [pinned to CPU 1] [pinned to CPU 1] [pinned to CPU 2] [pinned to CPU 3]
[normal priority] [idle priority] [normal priority] [normal priority] [normal priority]
busyloop
read(<pipe>) busyloop busyloop
y wa it ing
yield b reading procfs attempting to
stats write through
mremap(...) detect old mapping
PTE is moved wakeup mremap() [keep
write(<pipe>)
yield by preemption TLB
entry
sched_setaffinity so that we can block alive]
(task B, CPU 0) without yielding to B
keep
[preempted]
task B fallocate(...) free the pages
off the
CPU pread(<zygote reuse one of
code>) the pages as
page cache overwrite
the page
Confidential + Proprietary
 Android kernel exploit (app -> zygote)
task A task B task C task D task E
[pinned to CPU 0] [pinned to CPU 1] [pinned to CPU 1] [pinned to CPU 2] [pinned to CPU 3]
[normal priority] [idle priority] [normal priority] [normal priority] [normal priority]
busyloop
read(<pipe>) busyloop busyloop
y wa it ing
yield b reading procfs attempting to
stats write through
mremap(...) detect old mapping
PTE is moved wakeup mremap() [keep
write(<pipe>)
yield by preemption TLB
entry
sched_setaffinity so that we can block alive]
(task B, CPU 0) without yielding to B
keep
[preempted]
task B fallocate(...) free the pages
off the
CPU pread(<zygote reuse one of
code>) the pages as
page cache overwrite
the page
Confidential + Proprietary
 Android kernel exploit (app -> zygote)
task A task B task C task D task E
[pinned to CPU 0] [pinned to CPU 1] [pinned to CPU 1] [pinned to CPU 2] [pinned to CPU 3]
[normal priority] [idle priority] [normal priority] [normal priority] [normal priority]
busyloop
read(<pipe>) busyloop busyloop
y wa it ing
yield b reading procfs attempting to
stats write through
mremap(...) detect old mapping
PTE is moved wakeup mremap() [keep
write(<pipe>)
yield by preemption TLB
entry
sched_setaffinity so that we can block alive]
(task B, CPU 0) without yielding to B
keep
[preempted]
task B fallocate(...) free the pages
off the
CPU pread(<zygote reuse one of
code>) the pages as
page cache overwrite
the page
Confidential + Proprietary
 Android kernel exploit (app -> zygote)
task A task B task C task D task E
[pinned to CPU 0] [pinned to CPU 1] [pinned to CPU 1] [pinned to CPU 2] [pinned to CPU 3]
[normal priority] [idle priority] [normal priority] [normal priority] [normal priority]
busyloop
read(<pipe>) busyloop busyloop
y wa it ing
yield b reading procfs attempting to
stats write through
mremap(...) detect old mapping
PTE is moved wakeup mremap() [keep
write(<pipe>)
yield by preemption TLB
entry
sched_setaffinity so that we can block alive]
(task B, CPU 0) without yielding to B
keep
[preempted]
task B fallocate(...) free the pages
off the
CPU pread(<zygote reuse one of
code>) the pages as
page cache overwrite
the page
Confidential + Proprietary
Bug 2: refcount decrement on
struct file
(yes, the bug doesn't involve a race condition, but the exploit kinda does)

Confidential + Proprietary
userfaultfd and FUSE
● userfaultfd() and FUSE allow userspace to synchronously handle page faults
● => userspace can block arbitrarily at copy_from_user()/copy_to_user()
● userfaultfd() and FUSE are not exposed to unprivileged Android code

=> not applicable on Android, but relevant on desktop Linux

Confidential + Proprietary
 FUSE for exploiting struct file refcount
overdecrement in Linux 4.4
f = fdget(insn->imm);
map = __bpf_map_get(f);
bug from 2016 to illustrate FUSE-based if (IS_ERR(map)) {
verbose("fd %d is not pointing to valid
use-after-free exploitation bpf_map\n", insn->imm);
fdput(f);
return PTR_ERR(map);
● file reference acquired with fdget()
}
● error path accidentally called fdput()
struct bpf_map *__bpf_map_get(struct fd f)
twice {
● struct file freed prematurely if (!f.file)
return ERR_PTR(-EBADF);
● use-after-free if (f.file->f_op != &bpf_map_fops) {
● exploited on Ubuntu 16.04 fdput(f);
return ERR_PTR(-EINVAL);
}

return f.file->private_data;
}
crbug.com/project-zero/808 Confidential + Proprietary
 kcmp() for reliable UAF
● CONFIG_CHECKPOINT_RESTORE static long kptr_obfuscate(long v, int type)
● smaller/equal/greater comparison {
return (v ^ cookies[type][0]) *
between permuted kernel pointers cookies[type][1];
● intended for grouping same-object }

references in O(n log(n)) static int kcmp_ptr(void *v1, void *v2, enum
kcmp_type type)
● works on: {
○ struct file long t1, t2;
○ struct mm_struct
○ struct files_struct t1 = kptr_obfuscate((long)v1, type);
t2 = kptr_obfuscate((long)v2, type);
○ struct fs_struct
○ struct sighand_struct return (t1 < t2) | ((t1 > t2) << 1);
○ struct io_context }
○ struct sem_undo_list
● tag reuse oracle for Memory Tagging
Confidential + Proprietary
unless tag bits are ignored
 FUSE for exploiting struct file refcount
overdecrement in Linux 4.4
● create FUSE mapping ssize_t vfs_writev(struct file *file, const
struct iovec __user *vec, [...]) {
● open writable file (/dev/null) if (!(file->f_mode & FMODE_WRITE))
● start writev() with iov in FUSE return -EBADF;
[...]
mapping return do_readv_writev(WRITE, file, vec,
vlen, pos);
● write mode check passes }
● import_iovec() stalls on page fault static ssize_t do_readv_writev(int type,
● trigger bug to free the file struct file *file, const struct iovec __user
* uvector, unsigned long nr_segs, loff_t
● open /etc/crontab as read-only *pos) {
[...]
● verify that struct file was allocated at ret = import_iovec(type, uvector, nr_segs,
ARRAY_SIZE(iovstack), &iov, &iter);
the same address with kcmp() (else [...]
re-open /etc/crontab) if (iter_fn)
ret = do_iter_readv_writev(file, &iter,
● resolve FUSE page fault pos, iter_fn);
[...]
● writev() writes into /etc/crontab } Confidential + Proprietary
Bug 3: use of getpidcon()

Confidential + Proprietary
int getpidcon(pid_t pid, char **context)

● userspace daemons need to check peer SELinux contexts

● unix domain sockets: SO_PEERSEC
● Android binder: until recently no context name, only sender
PID

fd = open("/proc/$pid/attr/current", O_RDONLY)
read(fd, buf, len)

Confidential + Proprietary
Bug 3: race condition in hwservicemanager
crbug.com/project-zero/1741

● receive binder IPC call (with caller PID)

● getpidcon(pid, &context)
● ACL check for context

➢ exit and make privileged thread reuse the PID

● race window can be widened to ~15s
Confidential + Proprietary
i_mutex on kernel 4.4
● sys_getdents() (for readdir()) iterates directory entries and
copies to userspace under inode->i_mutex
○ potentially a large amount of data if the directory has many
entries
● lookup_slow() (for looking up uncached directory entries)
takes parent->d_inode->i_mutex
● => blocking userspace access in the middle of
sys_getdents() blocks concurrent path traversal (e.g.
open()) on the same inode

(Linux >=4.7 uses a semaphore i_rwsem in read mode instead of

i_mutex) Confidential + Proprietary
Priority Inversion
● high-priority task blocks on mutex task A block on lock
held by low-priority task (high priority)

● low-priority task is preempted by

medium-priority task (same CPU)
task B
long-running task
● also applies for violating fairness (normal priority)
between two normal-priority tasks
preemption
● kernel mutexes are vulnerable to
task C take
priority inversion! (low priority) lock
○ (unless you're on PREEMPT_RT)
● => we can artificially create a
priority inversion problem time
● mitigated by infrequent idle-priority
scheduling Confidential + Proprietary
 I/O

Major faults task A

[idle]
lock copy_to_user() #PF unlock
runnable?
yield no preempt resched
task B
spinning
Instead of userfaultfd(): [normal]

● create an uncached writable file mapping

○ [by filling up RAM with other data to force page cache eviction]
● let A trigger copy_to_user() on the file mapping while holding a lock
● let B spinloop at the same time

Consequences:

● copy_to_user() enters disk I/O path

● I/O path sleeps until disk responds, yielding the CPU
● scheduler won't preempt B when A is runnable again
Confidential + Proprietary
Repeated file mapping faults
● [map pages such that readahead logic can't fire]
● 83560 bytes output from sys_getdents() = 21 pages
[rounded up]
● >1s delay per disk read because of scheduler policy
● => >21s total delay
I/O I/O I/O
task A
lock copy_from_user() unlock
[idle]
runnable?
yield resched yield resched yield resched
task B
spinning
[normal]
Confidential + Proprietary
Confidential + Proprietary
Click to edit title
● Click to edit text
○ Second level
■ Third level
● Fourth level

○ Fifth level

Confidential + Proprietary
Click to place
text here

Confidential + Proprietary
 Timing diagram
task A task B task C task D (simplified)
[pinned to CPU 0] [pinned to CPU 0] [normal priority] [normal priority]
[normal priority] [IDLE priority]
busyloop open binder

keep
task B
off the
CPU

Confidential + Proprietary