The HSR test, hsr_ping.sh, actually needs 7 min to run. Around 375s to
be exact, and even more on a debug kernel or kernel with other network
security limits. The timeout setting for the kselftest is currently 45
seconds, which is way too short to integrate hsr tests to run_kselftest
infrastructure. However, timeout of hundreds of seconds is quite a long
time, especially in a CI/CD environment. It seems that we need
accelerate the test and balance with timeout setting.
The most time-consuming func is do_ping_long, where ping command sends
10 packages to the given address. The default interval between two ping
packages is 1s according to the ping Mannual. There isn't any operation
between pings thus we could pass -i 0.1 to ping to make it 10 times
faster.
While even with this short interval, the test still need about 46.4
seconds to finish because of the two HSR interfaces, each of which is
tested by calling do_ping func 12 times and do_ping_long func 19 times
and sleep for 3s.
So, an explicit setting is also needed to slightly increase the
timeout. And to leave us some slack, use 50 as default timeout.
Signed-off-by: Yunshui Jiang <jiangyunshui@kylinos.cn>
Link: https://patch.msgid.link/20241028082757.2945232-1-jiangyunshui@kylinos.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add more test cases for bits iterator:
(1) huge word test
Verify the multiplication overflow of nr_bits in bits_iter. Without
the overflow check, when nr_words is 67108865, nr_bits becomes 64,
causing bpf_probe_read_kernel_common() to corrupt the stack.
(2) max word test
Verify correct handling of maximum nr_words value (511).
(3) bad word test
Verify early termination of bits iteration when bits iterator
initialization fails.
Also rename bits_nomem to bits_too_big to better reflect its purpose.
Signed-off-by: Hou Tao <houtao1@huawei.com>
Link: https://lore.kernel.org/r/20241030100516.3633640-6-houtao@huaweicloud.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Add a simple file stressor that lives directly in-tree. This will create
a bunch of processes that each open 500 file descriptors and then use
close_range() to close them all.
Concurrently, other processes read /proc/<pid>/fd/ which rougly does
f = fget_task_next(p, &fd);
if (!f)
break;
data.mode = f->f_mode;
fput(f);
Which means that it'll try to get a reference to a file in another
task's file descriptor table.
Under heavy file load it is increasingly likely that the other task will
manage to close @file and @file will be recycled due to
SLAB_TYPEAFE_BY_RCU concurrently. This will trigger various warnings in
the file reference counting code.
Link: https://lore.kernel.org/r/20241021-vergab-streuen-924df15dceb9@brauner
Signed-off-by: Christian Brauner <brauner@kernel.org>
Pull sched_ext fixes from Tejun Heo:
- Instances of scx_ops_bypass() could race each other leading to
misbehavior. Fix by protecting the operation with a spinlock.
- selftest and userspace header fixes
* tag 'sched_ext-for-6.12-rc5-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext:
sched_ext: Fix enq_last_no_enq_fails selftest
sched_ext: Make cast_mask() inline
scx: Fix raciness in scx_ops_bypass()
scx: Fix exit selftest to use custom DSQ
sched_ext: Fix function pointer type mismatches in BPF selftests
selftests/sched_ext: add order-only dependency of runner.o on BPFOBJ
Drop bpf_iter.h header which uses vmlinux.h but re-defines a bunch of
iterator structures and some of BPF constants for use in BPF iterator
selftests.
None of that is necessary when fresh vmlinux.h header is generated for
vmlinux image that matches latest selftests. So drop ugly hacks and have
a nice plain vmlinux.h usage everywhere.
We could do the same with all the kfunc __ksym redefinitions, but that
has dependency on very fresh pahole, so I'm not addressing that here.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20241029203919.1948941-1-andrii@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The CI occasionaly encounters a failing test run. Example:
# PASS: ipsec tunnel mode for ns1/ns2
# re-run with random mtus: -o 10966 -l 19499 -r 31322
# PASS: flow offloaded for ns1/ns2
[..]
# FAIL: ipsec tunnel ... counter 1157059 exceeds expected value 878489
This script will re-exec itself, on the second run, random MTUs are
chosen for the involved links. This is done so we can cover different
combinations (large mtu on client, small on server, link has lowest
mtu, etc).
Furthermore, file size is random, even for the first run.
Rework this script and always use the same file size on initial run so
that at least the first round can be expected to have reproducible
behavior.
Second round will use random mtu/filesize.
Raise the failure limit to that of the file size, this should avoid all
errneous test errors. Currently, first fin will remove the offload, so if
one peer is already closing remaining data is handled by classic path,
which result in larger-than-expected counter and a test failure.
Given packet path also counts tcp/ip headers, in case offload is
completely broken this test will still fail (as expected).
The test counter limit could be made more strict again in the future
once flowtable can keep a connection in offloaded state until FINs
in both directions were seen.
Signed-off-by: Florian Westphal <fw@strlen.de>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20241022152324.13554-1-fw@strlen.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The test is motivated by the following observation:
Raise a signal, jump to signal handler. The ucontext_t structure dumped
by kernel to userspace has a uc_sigmask field having the mask of blocked
signals. If you run a fresh minimalistic program doing this, this field
is empty, even if you block some signals while registering the handler
with sigaction().
Here is what the man-pages have to say:
sigaction(2): "sa_mask specifies a mask of signals which should be blocked
(i.e., added to the signal mask of the thread in which the signal handler
is invoked) during execution of the signal handler. In addition, the
signal which triggered the handler will be blocked, unless the SA_NODEFER
flag is used."
signal(7): Under "Execution of signal handlers", (1.3) implies:
"The thread's current signal mask is accessible via the ucontext_t
object that is pointed to by the third argument of the signal handler."
But, (1.4) states:
"Any signals specified in act->sa_mask when registering the handler with
sigprocmask(2) are added to the thread's signal mask. The signal being
delivered is also added to the signal mask, unless SA_NODEFER was
specified when registering the handler. These signals are thus blocked
while the handler executes."
There clearly is no distinction being made in the man pages between
"Thread's signal mask" and ucontext_t; this logically should imply
that a signal blocked by populating struct sigaction should be visible
in ucontext_t.
Here is what the kernel code does (for Aarch64):
do_signal() -> handle_signal() -> sigmask_to_save(), which returns
¤t->blocked, is passed to setup_rt_frame() -> setup_sigframe() ->
__copy_to_user(). Hence, ¤t->blocked is copied to ucontext_t
exposed to userspace. Returning back to handle_signal(),
signal_setup_done() -> signal_delivered() -> sigorsets() and
set_current_blocked() are responsible for using information from
struct ksignal ksig, which was populated through the sigaction()
system call in kernel/signal.c:
copy_from_user(&new_sa.sa, act, sizeof(new_sa.sa)),
to update ¤t->blocked; hence, the set of blocked signals for the
current thread is updated AFTER the kernel dumps ucontext_t to
userspace.
Assuming that the above is indeed the intended behaviour, because it
semantically makes sense, since the signals blocked using sigaction()
remain blocked only till the execution of the handler, and not in the
context present before jumping to the handler (but nothing can be
confirmed from the man-pages), this patch introduces a test for
mangling with uc_sigmask.
The test asserts the relation between blocked signal, delivered signal,
and ucontext. The ucontext is mangled with, by adding a signal mask to
it; on return from the handler, the thread must block the corresponding
signal.
In the test description, I have also described signal delivery and blockage,
for ease of understanding what the test does.
Signed-off-by: Dev Jain <dev.jain@arm.com>
Reviewed-by: Mark Brown <broonie@kernel.org>
Signed-off-by: Shuah Khan <skhan@linuxfoundation.org>
Patch series "selftests/mm: revert pthread_barrier change"
On Android arm, pthread_create followed by a fork caused a deadlock in
the case where the fork required work to be completed by the created
thread.
The previous patches incorrectly assumed that the parent would
always initialize the pthread_barrier for the child thread. This
reverts the change and replaces the fix for wp-fork-with-event with the
original use of atomic_bool.
This patch (of 3):
This reverts commit e142cc87ac.
fork_event_consumer may be called by other tests that do not initialize
the pthread_barrier, so this approach is not correct. The subsequent
patch will revert to using atomic_bool instead.
Link: https://lkml.kernel.org/r/20241018171734.2315053-1-edliaw@google.com
Link: https://lkml.kernel.org/r/20241018171734.2315053-2-edliaw@google.com
Fixes: e142cc87ac ("fix deadlock for fork after pthread_create on ARM")
Signed-off-by: Edward Liaw <edliaw@google.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
When running watchdog-test with 'make run_tests', the watchdog-test will
be terminated by a timeout signal(SIGTERM) due to the test timemout.
And then, a system reboot would happen due to watchdog not stop. see
the dmesg as below:
```
[ 1367.185172] watchdog: watchdog0: watchdog did not stop!
```
Fix it by registering more signals(including SIGTERM) in watchdog-test,
where its signal handler will stop the watchdog.
After that
# timeout 1 ./watchdog-test
Watchdog Ticking Away!
.
Stopping watchdog ticks...
Link: https://lore.kernel.org/all/20241029031324.482800-1-lizhijian@fujitsu.com/
Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
Reviewed-by: Shuah Khan <skhan@linuxfoundation.org>
Signed-off-by: Shuah Khan <skhan@linuxfoundation.org>
Test case idmap_mount_tree_invalid failed to run on the newer kernel
with the following output:
# RUN mount_setattr_idmapped.idmap_mount_tree_invalid ...
# mount_setattr_test.c:1428:idmap_mount_tree_invalid:Expected sys_mount_setattr(open_tree_fd, "", AT_EMPTY_PATH, &attr, sizeof(attr)) (0) ! = 0 (0)
# idmap_mount_tree_invalid: Test terminated by assertion
This is because tmpfs is mounted at "/mnt/A", and tmpfs already
contains the flag FS_ALLOW_IDMAP after the commit 7a80e5b8c6 ("shmem:
support idmapped mounts for tmpfs"). So calling sys_mount_setattr here
returns 0 instead of -EINVAL as expected.
Ramfs does not support idmap mounts, so we can use it here to test invalid mounts,
which allows the test case to pass with the following output:
# Starting 1 tests from 1 test cases.
# RUN mount_setattr_idmapped.idmap_mount_tree_invalid ...
# OK mount_setattr_idmapped.idmap_mount_tree_invalid
ok 1 mount_setattr_idmapped.idmap_mount_tree_invalid
# PASSED: 1 / 1 tests passed.
Link: https://lore.kernel.org/all/20241028084132.3212598-1-zhouyuhang1010@163.com/
Signed-off-by: zhouyuhang <zhouyuhang@kylinos.cn>
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Shuah Khan <skhan@linuxfoundation.org>
Add tests that check if getsockopt(TCP_AO_GET_KEYS) returns the right
keys when using different filters.
Sample output:
> # ok 114 filter keys: by sndid, rcvid, address
> # ok 115 filter keys: by is_current
> # ok 116 filter keys: by is_rnext
> # ok 117 filter keys: by sndid, rcvid
> # ok 118 filter keys: correct nkeys when in.nkeys < matches
Acked-by: Dmitry Safonov <0x7f454c46@gmail.com>
Signed-off-by: Leo Stone <leocstone@gmail.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20241021174652.6949-1-leocstone@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Listing all the values linked to the MPTCP sysctl knobs was not
exercised in MPTCP test suite.
Let's do that to avoid any regressions, but also to have a kernel with a
debug kconfig verifying more assumptions. For the moment, we are not
interested by the output, only to avoid crashes and warnings.
Signed-off-by: Matthieu Baerts (NGI0) <matttbe@kernel.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20241021-net-mptcp-sched-lock-v1-3-637759cf061c@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Test that the plumbing exposed to userspace for injecting aborts in
response to unexpected MMIO works as intended in two different flavors:
- A 'normal' MMIO instruction (i.e. ESR_ELx.ISV=1)
- An ISV=0 MMIO instruction with/without KVM_CAP_ARM_NISV_TO_USER
enabled
Reviewed-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20241025203106.3529261-5-oliver.upton@linux.dev
Signed-off-by: Oliver Upton <oliver.upton@linux.dev>
cc9877fb76 ("sched_ext: Improve error reporting during loading") changed
how load failures are reported so that more error context can be
communicated. This breaks the enq_last_no_enq_fails test as attach no longer
fails. The scheduler is guaranteed to be ejected on attach completion with
full error information. Update enq_last_no_enq_fails so that it checks that
the scheduler is ejected using ops.exit().
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Vishal Chourasia <vishalc@linux.ibm.com>
Link: http://lkml.kernel.org/r/Zxknp7RAVNjmdJSc@linux.ibm.com
Fixes: cc9877fb76 ("sched_ext: Improve error reporting during loading")
In support of investigating an initialization failure report [1],
cxl_test was updated to register mock memory-devices after the mock
root-port/bus device had been registered. That led to cxl_test crashing
with a use-after-free bug with the following signature:
cxl_port_attach_region: cxl region3: cxl_host_bridge.0:port3 decoder3.0 add: mem0:decoder7.0 @ 0 next: cxl_switch_uport.0 nr_eps: 1 nr_targets: 1
cxl_port_attach_region: cxl region3: cxl_host_bridge.0:port3 decoder3.0 add: mem4:decoder14.0 @ 1 next: cxl_switch_uport.0 nr_eps: 2 nr_targets: 1
cxl_port_setup_targets: cxl region3: cxl_switch_uport.0:port6 target[0] = cxl_switch_dport.0 for mem0:decoder7.0 @ 0
1) cxl_port_setup_targets: cxl region3: cxl_switch_uport.0:port6 target[1] = cxl_switch_dport.4 for mem4:decoder14.0 @ 1
[..]
cxld_unregister: cxl decoder14.0:
cxl_region_decode_reset: cxl_region region3:
mock_decoder_reset: cxl_port port3: decoder3.0 reset
2) mock_decoder_reset: cxl_port port3: decoder3.0: out of order reset, expected decoder3.1
cxl_endpoint_decoder_release: cxl decoder14.0:
[..]
cxld_unregister: cxl decoder7.0:
3) cxl_region_decode_reset: cxl_region region3:
Oops: general protection fault, probably for non-canonical address 0x6b6b6b6b6b6b6bc3: 0000 [#1] PREEMPT SMP PTI
[..]
RIP: 0010:to_cxl_port+0x8/0x60 [cxl_core]
[..]
Call Trace:
<TASK>
cxl_region_decode_reset+0x69/0x190 [cxl_core]
cxl_region_detach+0xe8/0x210 [cxl_core]
cxl_decoder_kill_region+0x27/0x40 [cxl_core]
cxld_unregister+0x5d/0x60 [cxl_core]
At 1) a region has been established with 2 endpoint decoders (7.0 and
14.0). Those endpoints share a common switch-decoder in the topology
(3.0). At teardown, 2), decoder14.0 is the first to be removed and hits
the "out of order reset case" in the switch decoder. The effect though
is that region3 cleanup is aborted leaving it in-tact and
referencing decoder14.0. At 3) the second attempt to teardown region3
trips over the stale decoder14.0 object which has long since been
deleted.
The fix here is to recognize that the CXL specification places no
mandate on in-order shutdown of switch-decoders, the driver enforces
in-order allocation, and hardware enforces in-order commit. So, rather
than fail and leave objects dangling, always remove them.
In support of making cxl_region_decode_reset() always succeed,
cxl_region_invalidate_memregion() failures are turned into warnings.
Crashing the kernel is ok there since system integrity is at risk if
caches cannot be managed around physical address mutation events like
CXL region destruction.
A new device_for_each_child_reverse_from() is added to cleanup
port->commit_end after all dependent decoders have been disabled. In
other words if decoders are allocated 0->1->2 and disabled 1->2->0 then
port->commit_end only decrements from 2 after 2 has been disabled, and
it decrements all the way to zero since 1 was disabled previously.
Link: http://lore.kernel.org/20241004212504.1246-1-gourry@gourry.net [1]
Cc: stable@vger.kernel.org
Fixes: 176baefb2e ("cxl/hdm: Commit decoder state to hardware")
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Alison Schofield <alison.schofield@intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: Zijun Hu <quic_zijuhu@quicinc.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Link: https://patch.msgid.link/172964782781.81806.17902885593105284330.stgit@dwillia2-xfh.jf.intel.com
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
In commit 63fb3ec805 ("sched_ext: Allow only user DSQs for
scx_bpf_consume(), scx_bpf_dsq_nr_queued() and bpf_iter_scx_dsq_new()"), we
updated the consume path to only accept user DSQs, thus making it invalid
to consume SCX_DSQ_GLOBAL. This selftest was doing that, so let's create a
custom DSQ and use that instead. The test now passes:
[root@virtme-ng sched_ext]# ./runner -t exit
===== START =====
TEST: exit
DESCRIPTION: Verify we can cleanly exit a scheduler in multiple places
OUTPUT:
[ 12.387229] sched_ext: BPF scheduler "exit" enabled
[ 12.406064] sched_ext: BPF scheduler "exit" disabled (unregistered from BPF)
[ 12.453325] sched_ext: BPF scheduler "exit" enabled
[ 12.474064] sched_ext: BPF scheduler "exit" disabled (unregistered from BPF)
[ 12.515241] sched_ext: BPF scheduler "exit" enabled
[ 12.532064] sched_ext: BPF scheduler "exit" disabled (unregistered from BPF)
[ 12.592063] sched_ext: BPF scheduler "exit" disabled (unregistered from BPF)
[ 12.654063] sched_ext: BPF scheduler "exit" disabled (unregistered from BPF)
[ 12.715062] sched_ext: BPF scheduler "exit" disabled (unregistered from BPF)
ok 1 exit #
===== END =====
Signed-off-by: David Vernet <void@manifault.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull bpf fixes from Daniel Borkmann:
- Fix an out-of-bounds read in bpf_link_show_fdinfo for BPF sockmap
link file descriptors (Hou Tao)
- Fix BPF arm64 JIT's address emission with tag-based KASAN enabled
reserving not enough size (Peter Collingbourne)
- Fix BPF verifier do_misc_fixups patching for inlining of the
bpf_get_branch_snapshot BPF helper (Andrii Nakryiko)
- Fix a BPF verifier bug and reject BPF program write attempts into
read-only marked BPF maps (Daniel Borkmann)
- Fix perf_event_detach_bpf_prog error handling by removing an invalid
check which would skip BPF program release (Jiri Olsa)
- Fix memory leak when parsing mount options for the BPF filesystem
(Hou Tao)
* tag 'bpf-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf:
bpf: Check validity of link->type in bpf_link_show_fdinfo()
bpf: Add the missing BPF_LINK_TYPE invocation for sockmap
bpf: fix do_misc_fixups() for bpf_get_branch_snapshot()
bpf,perf: Fix perf_event_detach_bpf_prog error handling
selftests/bpf: Add test for passing in uninit mtu_len
selftests/bpf: Add test for writes to .rodata
bpf: Remove MEM_UNINIT from skb/xdp MTU helpers
bpf: Fix overloading of MEM_UNINIT's meaning
bpf: Add MEM_WRITE attribute
bpf: Preserve param->string when parsing mount options
bpf, arm64: Fix address emission with tag-based KASAN enabled
This patch test the following failures in syscall update_elem
1. The first update_elem(BPF_F_LOCK) should be EOPNOTSUPP. syscall.c takes
care of unpinning the uptr.
2. The second update_elem(BPF_EXIST) fails. syscall.c takes care of
unpinning the uptr.
3. The forth update_elem(BPF_NOEXIST) fails. bpf_local_storage_update
takes care of unpinning the uptr.
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Link: https://lore.kernel.org/r/20241023234759.860539-11-martin.lau@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
This patch tests the case when uptr has a struct spanning across two
pages. It is not supported now and EOPNOTSUPP is expected from the
syscall update_elem.
It also tests the whole uptr struct located exactly at the
end of a page and ensures that this case is accepted by update_elem.
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Link: https://lore.kernel.org/r/20241023234759.860539-10-martin.lau@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
We lack find_symbol() selftests, so add one. This let's us stress test
improvements easily on find_symbol() or optimizations. It also inherently
allows us to test the limits of kallsyms on Linux today.
We test a pathalogical use case for kallsyms by introducing modules
which are automatically written for us with a larger number of symbols.
We have 4 kallsyms test modules:
A: has KALLSYSMS_NUMSYMS exported symbols
B: uses one of A's symbols
C: adds KALLSYMS_SCALE_FACTOR * KALLSYSMS_NUMSYMS exported
D: adds 2 * the symbols than C
By using anything much larger than KALLSYSMS_NUMSYMS as 10,000 and
KALLSYMS_SCALE_FACTOR of 8 we segfault today. So we're capped at
around 160000 symbols somehow today. We can inpsect that issue at
our leasure later, but for now the real value to this test is that
this will easily allow us to test improvements on find_symbol().
We want to enable this test on allyesmodconfig builds so we can't
use this combination, so instead just use a safe value for now and
be informative on the Kconfig symbol documentation about where our
thresholds are for testers. We default then to KALLSYSMS_NUMSYMS of
just 100 and KALLSYMS_SCALE_FACTOR of 8.
On x86_64 we can use perf, for other architectures we just use 'time'
and allow for customizations. For example a future enhancements could
be done for parisc to check for unaligned accesses which triggers a
special special exception handler assembler code inside the kernel.
The negative impact on performance is so large on parisc that it
keeps track of its accesses on /proc/cpuinfo as UAH:
IRQ: CPU0 CPU1
3: 1332 0 SuperIO ttyS0
7: 1270013 0 SuperIO pata_ns87415
64: 320023012 320021431 CPU timer
65: 17080507 20624423 CPU IPI
UAH: 10948640 58104 Unaligned access handler traps
While at it, this tidies up lib/ test modules to allow us to have
a new directory for them. The amount of test modules under lib/
is insane.
This should also hopefully showcase how to start doing basic
self module writing code, which may be more useful for more complex
cases later in the future.
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Fix incompatible function pointer type warnings in sched_ext BPF selftests by
explicitly casting the function pointers when initializing struct_ops.
This addresses multiple -Wincompatible-function-pointer-types warnings from the
clang compiler where function signatures didn't match exactly.
The void * cast ensures the compiler accepts the function pointer
assignment despite minor type differences in the parameters.
Signed-off-by: Vishal Chourasia <vishalc@linux.ibm.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
