* [PATCH bpf] bpf/tracing: fix a deadlock in perf_event_detach_bpf_prog
@ 2018-04-09 16:18 Yonghong Song
2018-04-09 16:47 ` Alexei Starovoitov
0 siblings, 1 reply; 4+ messages in thread
From: Yonghong Song @ 2018-04-09 16:18 UTC (permalink / raw)
To: ast, daniel, netdev; +Cc: kernel-team
syzbot reported a possible deadlock in perf_event_detach_bpf_prog.
The error details:
======================================================
WARNING: possible circular locking dependency detected
4.16.0-rc7+ #3 Not tainted
------------------------------------------------------
syz-executor7/24531 is trying to acquire lock:
(bpf_event_mutex){+.+.}, at: [<000000008a849b07>] perf_event_detach_bpf_prog+0x92/0x3d0 kernel/trace/bpf_trace.c:854
but task is already holding lock:
(&mm->mmap_sem){++++}, at: [<0000000038768f87>] vm_mmap_pgoff+0x198/0x280 mm/util.c:353
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #1 (&mm->mmap_sem){++++}:
__might_fault+0x13a/0x1d0 mm/memory.c:4571
_copy_to_user+0x2c/0xc0 lib/usercopy.c:25
copy_to_user include/linux/uaccess.h:155 [inline]
bpf_prog_array_copy_info+0xf2/0x1c0 kernel/bpf/core.c:1694
perf_event_query_prog_array+0x1c7/0x2c0 kernel/trace/bpf_trace.c:891
_perf_ioctl kernel/events/core.c:4750 [inline]
perf_ioctl+0x3e1/0x1480 kernel/events/core.c:4770
vfs_ioctl fs/ioctl.c:46 [inline]
do_vfs_ioctl+0x1b1/0x1520 fs/ioctl.c:686
SYSC_ioctl fs/ioctl.c:701 [inline]
SyS_ioctl+0x8f/0xc0 fs/ioctl.c:692
do_syscall_64+0x281/0x940 arch/x86/entry/common.c:287
entry_SYSCALL_64_after_hwframe+0x42/0xb7
-> #0 (bpf_event_mutex){+.+.}:
lock_acquire+0x1d5/0x580 kernel/locking/lockdep.c:3920
__mutex_lock_common kernel/locking/mutex.c:756 [inline]
__mutex_lock+0x16f/0x1a80 kernel/locking/mutex.c:893
mutex_lock_nested+0x16/0x20 kernel/locking/mutex.c:908
perf_event_detach_bpf_prog+0x92/0x3d0 kernel/trace/bpf_trace.c:854
perf_event_free_bpf_prog kernel/events/core.c:8147 [inline]
_free_event+0xbdb/0x10f0 kernel/events/core.c:4116
put_event+0x24/0x30 kernel/events/core.c:4204
perf_mmap_close+0x60d/0x1010 kernel/events/core.c:5172
remove_vma+0xb4/0x1b0 mm/mmap.c:172
remove_vma_list mm/mmap.c:2490 [inline]
do_munmap+0x82a/0xdf0 mm/mmap.c:2731
mmap_region+0x59e/0x15a0 mm/mmap.c:1646
do_mmap+0x6c0/0xe00 mm/mmap.c:1483
do_mmap_pgoff include/linux/mm.h:2223 [inline]
vm_mmap_pgoff+0x1de/0x280 mm/util.c:355
SYSC_mmap_pgoff mm/mmap.c:1533 [inline]
SyS_mmap_pgoff+0x462/0x5f0 mm/mmap.c:1491
SYSC_mmap arch/x86/kernel/sys_x86_64.c:100 [inline]
SyS_mmap+0x16/0x20 arch/x86/kernel/sys_x86_64.c:91
do_syscall_64+0x281/0x940 arch/x86/entry/common.c:287
entry_SYSCALL_64_after_hwframe+0x42/0xb7
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(&mm->mmap_sem);
lock(bpf_event_mutex);
lock(&mm->mmap_sem);
lock(bpf_event_mutex);
*** DEADLOCK ***
======================================================
The bug is introduced by Commit f371b304f12e ("bpf/tracing: allow
user space to query prog array on the same tp") where copy_to_user,
which requires mm->mmap_sem, is called inside bpf_event_mutex lock.
At the same time, during perf_event file descriptor close,
mm->mmap_sem is held first and then subsequent
perf_event_detach_bpf_prog needs bpf_event_mutex lock.
Such a senario caused a deadlock.
As suggested by Danial, moving copy_to_user out of the
bpf_event_mutex lock should fix the problem.
Fixes: f371b304f12e ("bpf/tracing: allow user space to query prog array on the same tp")
Reported-by: syzbot+dc5ca0e4c9bfafaf2bae@syzkaller.appspotmail.com
Signed-off-by: Yonghong Song <yhs@fb.com>
---
include/linux/bpf.h | 4 ++--
kernel/bpf/core.c | 25 +++++++++++++++++++------
kernel/trace/bpf_trace.c | 24 ++++++++++++++++++++----
3 files changed, 41 insertions(+), 12 deletions(-)
diff --git a/include/linux/bpf.h b/include/linux/bpf.h
index 95a7abd..486e65e 100644
--- a/include/linux/bpf.h
+++ b/include/linux/bpf.h
@@ -339,8 +339,8 @@ int bpf_prog_array_copy_to_user(struct bpf_prog_array __rcu *progs,
void bpf_prog_array_delete_safe(struct bpf_prog_array __rcu *progs,
struct bpf_prog *old_prog);
int bpf_prog_array_copy_info(struct bpf_prog_array __rcu *array,
- __u32 __user *prog_ids, u32 request_cnt,
- __u32 __user *prog_cnt);
+ u32 *prog_ids, u32 request_cnt,
+ u32 *prog_cnt);
int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array,
struct bpf_prog *exclude_prog,
struct bpf_prog *include_prog,
diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index d315b39..a95a7de 100644
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c
@@ -1683,22 +1683,35 @@ int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array,
}
int bpf_prog_array_copy_info(struct bpf_prog_array __rcu *array,
- __u32 __user *prog_ids, u32 request_cnt,
- __u32 __user *prog_cnt)
+ u32 *prog_ids, u32 request_cnt,
+ u32 *prog_cnt)
{
- u32 cnt = 0;
+ struct bpf_prog **prog;
+ u32 i = 0, cnt = 0;
if (array)
cnt = bpf_prog_array_length(array);
- if (copy_to_user(prog_cnt, &cnt, sizeof(cnt)))
- return -EFAULT;
+ *prog_cnt = cnt;
/* return early if user requested only program count or nothing to copy */
if (!request_cnt || !cnt)
return 0;
- return bpf_prog_array_copy_to_user(array, prog_ids, request_cnt);
+ /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
+ prog = rcu_dereference_check(array, 1)->progs;
+ for (; *prog; prog++) {
+ if (*prog == &dummy_bpf_prog.prog)
+ continue;
+ prog_ids[i] = (*prog)->aux->id;
+ if (++i == request_cnt) {
+ prog++;
+ break;
+ }
+ }
+ if (!!(*prog))
+ return -ENOSPC;
+ return 0;
}
static void bpf_prog_free_deferred(struct work_struct *work)
diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index d88e96d..bd1e458 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -977,6 +977,7 @@ int perf_event_query_prog_array(struct perf_event *event, void __user *info)
{
struct perf_event_query_bpf __user *uquery = info;
struct perf_event_query_bpf query = {};
+ u32 *ids, prog_cnt, ids_len;
int ret;
if (!capable(CAP_SYS_ADMIN))
@@ -985,16 +986,31 @@ int perf_event_query_prog_array(struct perf_event *event, void __user *info)
return -EINVAL;
if (copy_from_user(&query, uquery, sizeof(query)))
return -EFAULT;
- if (query.ids_len > BPF_TRACE_MAX_PROGS)
+
+ ids_len = query.ids_len;
+ if (ids_len > BPF_TRACE_MAX_PROGS)
return -E2BIG;
+ ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN);
+ if (!ids)
+ return -ENOMEM;
mutex_lock(&bpf_event_mutex);
ret = bpf_prog_array_copy_info(event->tp_event->prog_array,
- uquery->ids,
- query.ids_len,
- &uquery->prog_cnt);
+ ids,
+ ids_len,
+ &prog_cnt);
mutex_unlock(&bpf_event_mutex);
+ if (!ret || ret == -ENOSPC) {
+ if (copy_to_user(&uquery->prog_cnt, &prog_cnt, sizeof(prog_cnt)) ||
+ copy_to_user(uquery->ids, ids, ids_len * sizeof(u32))) {
+ ret = -EFAULT;
+ goto out;
+ }
+ }
+
+out:
+ kfree(ids);
return ret;
}
--
2.9.5
^ permalink raw reply related [flat|nested] 4+ messages in thread
* Re: [PATCH bpf] bpf/tracing: fix a deadlock in perf_event_detach_bpf_prog
2018-04-09 16:18 [PATCH bpf] bpf/tracing: fix a deadlock in perf_event_detach_bpf_prog Yonghong Song
@ 2018-04-09 16:47 ` Alexei Starovoitov
2018-04-09 18:41 ` Yonghong Song
0 siblings, 1 reply; 4+ messages in thread
From: Alexei Starovoitov @ 2018-04-09 16:47 UTC (permalink / raw)
To: Yonghong Song, daniel, netdev; +Cc: kernel-team
On 4/9/18 9:18 AM, Yonghong Song wrote:
> syzbot reported a possible deadlock in perf_event_detach_bpf_prog.
...
> @@ -985,16 +986,31 @@ int perf_event_query_prog_array(struct perf_event *event, void __user *info)
> return -EINVAL;
> if (copy_from_user(&query, uquery, sizeof(query)))
> return -EFAULT;
> - if (query.ids_len > BPF_TRACE_MAX_PROGS)
> +
> + ids_len = query.ids_len;
> + if (ids_len > BPF_TRACE_MAX_PROGS)
> return -E2BIG;
> + ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN);
> + if (!ids)
> + return -ENOMEM;
>
> mutex_lock(&bpf_event_mutex);
> ret = bpf_prog_array_copy_info(event->tp_event->prog_array,
> - uquery->ids,
> - query.ids_len,
> - &uquery->prog_cnt);
> + ids,
> + ids_len,
> + &prog_cnt);
> mutex_unlock(&bpf_event_mutex);
>
> + if (!ret || ret == -ENOSPC) {
> + if (copy_to_user(&uquery->prog_cnt, &prog_cnt, sizeof(prog_cnt)) ||
> + copy_to_user(uquery->ids, ids, ids_len * sizeof(u32))) {
> + ret = -EFAULT;
> + goto out;
> + }
> + }
> +
> +out:
> + kfree(ids);
alloc/free just to avoid this locking dependency feels suboptimal.
may be we can get rid of bpf_event_mutex in some cases?
the perf event itself is locked via perf_event_ctx_lock() when we're
calling perf_event_query_prog_array, perf_event_attach|detach_bpf_prog.
I forgot what was the motivation for us to introduce it in the
first place.
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH bpf] bpf/tracing: fix a deadlock in perf_event_detach_bpf_prog
2018-04-09 16:47 ` Alexei Starovoitov
@ 2018-04-09 18:41 ` Yonghong Song
2018-04-10 0:51 ` Alexei Starovoitov
0 siblings, 1 reply; 4+ messages in thread
From: Yonghong Song @ 2018-04-09 18:41 UTC (permalink / raw)
To: Alexei Starovoitov, daniel, netdev; +Cc: kernel-team
On 4/9/18 9:47 AM, Alexei Starovoitov wrote:
> On 4/9/18 9:18 AM, Yonghong Song wrote:
>> syzbot reported a possible deadlock in perf_event_detach_bpf_prog.
> ...
>> @@ -985,16 +986,31 @@ int perf_event_query_prog_array(struct
>> perf_event *event, void __user *info)
>> return -EINVAL;
>> if (copy_from_user(&query, uquery, sizeof(query)))
>> return -EFAULT;
>> - if (query.ids_len > BPF_TRACE_MAX_PROGS)
>> +
>> + ids_len = query.ids_len;
>> + if (ids_len > BPF_TRACE_MAX_PROGS)
>> return -E2BIG;
>> + ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN);
>> + if (!ids)
>> + return -ENOMEM;
>>
>> mutex_lock(&bpf_event_mutex);
>> ret = bpf_prog_array_copy_info(event->tp_event->prog_array,
>> - uquery->ids,
>> - query.ids_len,
>> - &uquery->prog_cnt);
>> + ids,
>> + ids_len,
>> + &prog_cnt);
>> mutex_unlock(&bpf_event_mutex);
>>
>> + if (!ret || ret == -ENOSPC) {
>> + if (copy_to_user(&uquery->prog_cnt, &prog_cnt,
>> sizeof(prog_cnt)) ||
>> + copy_to_user(uquery->ids, ids, ids_len * sizeof(u32))) {
>> + ret = -EFAULT;
>> + goto out;
>> + }
>> + }
>> +
>> +out:
>> + kfree(ids);
>
> alloc/free just to avoid this locking dependency feels suboptimal.
We actually already did kcalloc/kfree in bpf_prog_array_copy_to_user.
In that function, we did not copy_to_user one id at a time.
We allocate a temporary array and store the result there
and at the end, we call one copy_to_user to copy to the user buffer.
The patch here just moved this allocation and associated copy_to_user
out of the function and bpf_event_mutex. It did not introduce new
allocations.
>
> may be we can get rid of bpf_event_mutex in some cases?
> the perf event itself is locked via perf_event_ctx_lock() when we're
> calling perf_event_query_prog_array, perf_event_attach|detach_bpf_prog.
> I forgot what was the motivation for us to introduce it in the
> first place.
The original motivation for the lock to make sure bpf_prog_array
does not change during middle of attach/detach/query. it looks like
we have:
. perf_event_attach|query under perf_event_ctx_lock
. perf_event_detach not under perf_event_ctx_lock
Introducing perf_event_ctx_lock to perf_event_detach could still
have the deadlock.
>
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH bpf] bpf/tracing: fix a deadlock in perf_event_detach_bpf_prog
2018-04-09 18:41 ` Yonghong Song
@ 2018-04-10 0:51 ` Alexei Starovoitov
0 siblings, 0 replies; 4+ messages in thread
From: Alexei Starovoitov @ 2018-04-10 0:51 UTC (permalink / raw)
To: Yonghong Song, daniel, netdev; +Cc: kernel-team
On 4/9/18 11:41 AM, Yonghong Song wrote:
>
>
> On 4/9/18 9:47 AM, Alexei Starovoitov wrote:
>> On 4/9/18 9:18 AM, Yonghong Song wrote:
>>> syzbot reported a possible deadlock in perf_event_detach_bpf_prog.
>> ...
>>> @@ -985,16 +986,31 @@ int perf_event_query_prog_array(struct
>>> perf_event *event, void __user *info)
>>> return -EINVAL;
>>> if (copy_from_user(&query, uquery, sizeof(query)))
>>> return -EFAULT;
>>> - if (query.ids_len > BPF_TRACE_MAX_PROGS)
>>> +
>>> + ids_len = query.ids_len;
>>> + if (ids_len > BPF_TRACE_MAX_PROGS)
>>> return -E2BIG;
>>> + ids = kcalloc(ids_len, sizeof(u32), GFP_USER | __GFP_NOWARN);
>>> + if (!ids)
>>> + return -ENOMEM;
>>>
>>> mutex_lock(&bpf_event_mutex);
>>> ret = bpf_prog_array_copy_info(event->tp_event->prog_array,
>>> - uquery->ids,
>>> - query.ids_len,
>>> - &uquery->prog_cnt);
>>> + ids,
>>> + ids_len,
>>> + &prog_cnt);
>>> mutex_unlock(&bpf_event_mutex);
>>>
>>> + if (!ret || ret == -ENOSPC) {
>>> + if (copy_to_user(&uquery->prog_cnt, &prog_cnt,
>>> sizeof(prog_cnt)) ||
>>> + copy_to_user(uquery->ids, ids, ids_len * sizeof(u32))) {
>>> + ret = -EFAULT;
>>> + goto out;
>>> + }
>>> + }
>>> +
>>> +out:
>>> + kfree(ids);
>>
>> alloc/free just to avoid this locking dependency feels suboptimal.
>
> We actually already did kcalloc/kfree in bpf_prog_array_copy_to_user.
> In that function, we did not copy_to_user one id at a time.
> We allocate a temporary array and store the result there
> and at the end, we call one copy_to_user to copy to the user buffer.
>
> The patch here just moved this allocation and associated copy_to_user
> out of the function and bpf_event_mutex. It did not introduce new
> allocations.
I see, so the patch essentially open coding
bpf_prog_array_copy_to_user()
can we share the code then?
bpf/core.c callsite used by trace/bpf_trace.c
and similar callsite in bpf/cgroup.c
should be using common helper.
>>
>> may be we can get rid of bpf_event_mutex in some cases?
>> the perf event itself is locked via perf_event_ctx_lock() when we're
>> calling perf_event_query_prog_array, perf_event_attach|detach_bpf_prog.
>> I forgot what was the motivation for us to introduce it in the
>> first place.
>
> The original motivation for the lock to make sure bpf_prog_array
> does not change during middle of attach/detach/query. it looks like
> we have:
> . perf_event_attach|query under perf_event_ctx_lock
> . perf_event_detach not under perf_event_ctx_lock
> Introducing perf_event_ctx_lock to perf_event_detach could still
> have the deadlock.
ahh, right, since the progs are in even->tp_event which can
be shared by multiple perf_events.
Scratch that idea.
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2018-04-10 0:51 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2018-04-09 16:18 [PATCH bpf] bpf/tracing: fix a deadlock in perf_event_detach_bpf_prog Yonghong Song
2018-04-09 16:47 ` Alexei Starovoitov
2018-04-09 18:41 ` Yonghong Song
2018-04-10 0:51 ` Alexei Starovoitov
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.