If the kmem_cache refcount is greater than one, we should not mark the
root kmem_cache as dying. If we mark the root kmem_cache dying
incorrectly, the non-root kmem_cache can never be destroyed. It
resulted in memory leak when memcg was destroyed. We can use the
following steps to reproduce.
1) Use kmem_cache_create() to create a new kmem_cache named A.
2) Coincidentally, the kmem_cache A is an alias for kmem_cache B,
so the refcount of B is just increased.
3) Use kmem_cache_destroy() to destroy the kmem_cache A, just
decrease the B's refcount but mark the B as dying.
4) Create a new memory cgroup and alloc memory from the kmem_cache
B. It leads to create a non-root kmem_cache for allocating memory.
5) When destroy the memory cgroup created in the step 4), the
non-root kmem_cache can never be destroyed.
If we repeat steps 4) and 5), this will cause a lot of memory leak. So
only when refcount reach zero, we mark the root kmem_cache as dying.
Fixes: 92ee383f6d ("mm: fix race between kmem_cache destroy, create and deactivate")
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/20200716165103.83462-1-songmuchun@bytedance.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It was hard to keep a test running, moving tasks between memcgs with
move_charge_at_immigrate, while swapping: mem_cgroup_id_get_many()'s
refcount is discovered to be 0 (supposedly impossible), so it is then
forced to REFCOUNT_SATURATED, and after thousands of warnings in quick
succession, the test is at last put out of misery by being OOM killed.
This is because of the way moved_swap accounting was saved up until the
task move gets completed in __mem_cgroup_clear_mc(), deferred from when
mem_cgroup_move_swap_account() actually exchanged old and new ids.
Concurrent activity can free up swap quicker than the task is scanned,
bringing id refcount down 0 (which should only be possible when
offlining).
Just skip that optimization: do that part of the accounting immediately.
Fixes: 615d66c37c ("mm: memcontrol: fix memcg id ref counter on swap charge move")
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/alpine.LSU.2.11.2007071431050.4726@eggly.anvils
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Prabhakar reported an OOPS inside mem_cgroup_get_nr_swap_pages()
function in a corner case seen on some arm64 boards when kdump kernel
runs with "cgroup_disable=memory" passed to the kdump kernel via
bootargs.
The root-cause behind the same is that currently mem_cgroup_swap_init()
function is implemented as a subsys_initcall() call instead of a
core_initcall(), this means 'cgroup_memory_noswap' still remains set to
the default value (false) even when memcg is disabled via
"cgroup_disable=memory" boot parameter.
This may result in premature OOPS inside mem_cgroup_get_nr_swap_pages()
function in corner cases:
Unable to handle kernel NULL pointer dereference at virtual address 0000000000000188
Mem abort info:
ESR = 0x96000006
EC = 0x25: DABT (current EL), IL = 32 bits
SET = 0, FnV = 0
EA = 0, S1PTW = 0
Data abort info:
ISV = 0, ISS = 0x00000006
CM = 0, WnR = 0
[0000000000000188] user address but active_mm is swapper
Internal error: Oops: 96000006 [#1] SMP
Modules linked in:
<..snip..>
Call trace:
mem_cgroup_get_nr_swap_pages+0x9c/0xf4
shrink_lruvec+0x404/0x4f8
shrink_node+0x1a8/0x688
do_try_to_free_pages+0xe8/0x448
try_to_free_pages+0x110/0x230
__alloc_pages_slowpath.constprop.106+0x2b8/0xb48
__alloc_pages_nodemask+0x2ac/0x2f8
alloc_page_interleave+0x20/0x90
alloc_pages_current+0xdc/0xf8
atomic_pool_expand+0x60/0x210
__dma_atomic_pool_init+0x50/0xa4
dma_atomic_pool_init+0xac/0x158
do_one_initcall+0x50/0x218
kernel_init_freeable+0x22c/0x2d0
kernel_init+0x18/0x110
ret_from_fork+0x10/0x18
Code: aa1403e3 91106000 97f82a27 14000011 (f940c663)
---[ end trace 9795948475817de4 ]---
Kernel panic - not syncing: Fatal exception
Rebooting in 10 seconds..
Fixes: eccb52e788 ("mm: memcontrol: prepare swap controller setup for integration")
Reported-by: Prabhakar Kushwaha <pkushwaha@marvell.com>
Signed-off-by: Bhupesh Sharma <bhsharma@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: James Morse <james.morse@arm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Link: http://lkml.kernel.org/r/1593641660-13254-2-git-send-email-bhsharma@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
clang static analysis reports a garbage return
In file included from mm/memory.c:84:
mm/memory.c:1612:2: warning: Undefined or garbage value returned to caller [core.uninitialized.UndefReturn]
return err;
^~~~~~~~~~
The setting of err depends on a loop executing. So initialize err.
Signed-off-by: Tom Rix <trix@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/20200703155354.29132-1-trix@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
After commit fdc85222d5 ("kernfs: kvmalloc xattr value instead of
kmalloc"), simple xattr entry is allocated with kvmalloc() instead of
kmalloc(), so we should release it with kvfree() instead of kfree().
Fixes: fdc85222d5 ("kernfs: kvmalloc xattr value instead of kmalloc")
Signed-off-by: Chengguang Xu <cgxu519@mykernel.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Tejun Heo <tj@kernel.org>
Cc: Daniel Xu <dxu@dxuuu.xyz>
Cc: Chris Down <chris@chrisdown.name>
Cc: Andreas Dilger <adilger@dilger.ca>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: <stable@vger.kernel.org> [5.7]
Link: http://lkml.kernel.org/r/20200704051608.15043-1-cgxu519@mykernel.net
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
VMA with VM_GROWSDOWN or VM_GROWSUP flag set can change their size under
mmap_read_lock(). It can lead to race with __do_munmap():
Thread A Thread B
__do_munmap()
detach_vmas_to_be_unmapped()
mmap_write_downgrade()
expand_downwards()
vma->vm_start = address;
// The VMA now overlaps with
// VMAs detached by the Thread A
// page fault populates expanded part
// of the VMA
unmap_region()
// Zaps pagetables partly
// populated by Thread B
Similar race exists for expand_upwards().
The fix is to avoid downgrading mmap_lock in __do_munmap() if detached
VMAs are next to VM_GROWSDOWN or VM_GROWSUP VMA.
[akpm@linux-foundation.org: s/mmap_sem/mmap_lock/ in comment]
Fixes: dd2283f260 ("mm: mmap: zap pages with read mmap_sem in munmap")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Yang Shi <yang.shi@linux.alibaba.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: <stable@vger.kernel.org> [4.20+]
Link: http://lkml.kernel.org/r/20200709105309.42495-1-kirill.shutemov@linux.intel.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Ports should be defined in the devicetree if they are to be enabled on
the system.
Signed-off-by: Eddie James <eajames@linux.ibm.com>
Signed-off-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: Wolfram Sang <wsa@kernel.org>
Up to ConnectX-7 setting mkey relaxed ordering read/write attributes
by UMR is not supported. ConnectX-7 supports this option, which is
indicated by two new HCA capabilities - relaxed_ordering_write_umr
and relaxed_ordering_read_umr.
Signed-off-by: Meir Lichtinger <meirl@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
The automatic object lifetime model allows us to change the write()
interface to have the same logic as the ioctl() path. Update the
create/alloc functions to be in the following format, so the code flow
will be the same:
* Allocate objects
* Initialize them
* Call to the drivers, this is last step that is allowed to fail
* Finalize object
* Return response and allow to core code to handle abort/commit
respectively.
Link: https://lore.kernel.org/r/20200719052223.75245-3-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Move io_req_init_async() into io_grab_files(), it's safer this way. Note
that io_queue_async_work() does *init_async(), so it's valid to move out
of __io_queue_sqe() punt path. Also, add a helper around io_grab_files().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Calling into opcode prep handlers may be dangerous, as they re-read
SQE but might not re-initialise requests completely. If io_req_defer()
passed fast checks and is done with preparations, punt it async.
As all other cases are covered with nulling @sqe, this guarantees that
io_[opcode]_prep() are visited only once per request.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In io_sq_thread(), if there are task works to handle, current codes
will skip schedule() and go on polling sq again, but forget to clear
IORING_SQ_NEED_WAKEUP flag, fix this issue. Also add two helpers to
set and clear IORING_SQ_NEED_WAKEUP flag,
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As every iopoll request have a task ref, it becomes expensive to put
them one by one, instead we can put several at once integrating that
into io_req_free_batch().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
put_task_struct_many() is as put_task_struct() but puts several
references at once. Useful to batching it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Locked and pinned memory accounting in io_{,un}account_mem() depends on
having ->sqo_mm, which is NULL after a recent change for non SQPOLL'ed
io_ring. That disables the accounting.
Return ->sqo_mm initialisation back, and do __io_sq_thread_acquire_mm()
based on IORING_SETUP_SQPOLL flag.
Fixes: 8eb06d7e8d ("io_uring: fix missing ->mm on exit")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_sqe_buffer_unregister() uses cxt->sqo_mm for memory accounting, but
io_ring_ctx_free() drops ->sqo_mm before leaving pinned_vm
over-accounted. Postpone mm cleanup for when it's not needed anymore.
Fixes: 309758254e ("io_uring: report pinned memory usage")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Don't implement fast path of kbuf freeing and management inlined into
io_recv{,msg}(), that's error prone and duplicates handling. Replace it
with a helper io_put_recv_kbuf(), which mimics io_put_rw_kbuf() in the
io_read/write().
This also keeps cflags calculation in one place, removing duplication
between rw and recv/send.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Extract a common helper for cleaning up a selected buffer, this will be
used shortly. By the way, correct cflags types to unsigned and, as kbufs
are anyway tracked by a flag, remove useless zeroing req->rw.addr.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move REQ_F_BUFFER_SELECT flag check out of io_recv_buffer_select(), and
do that in its call sites That saves us from double error checking and
possibly an extra function call.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_clean_op() may be skipped even if there is a selected io_buffer,
that's because *select_buffer() funcions never set REQ_F_NEED_CLEANUP.
Trigger io_clean_op() when REQ_F_BUFFER_SELECTED is set as well, and
and clear the flag if was freed out of it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Instead of returning error from io_recv(), go through generic cleanup
path, because it'll retain cflags for userspace. Do the same for
io_send() for consistency.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With the return on a bad socket, kmsg is always non-null by the end
of the function, prune left extra checks and initialisations.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Flip over "if (sock)" condition with return on error, the upper layer
will take care. That change will be handy later, but already removes
an extra jump from hot path.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, file refs in struct io_submit_state are tracked with 2 vars:
@has_refs -- how many refs were initially taken
@used_refs -- number of refs used
Replace it with a single variable counting how many refs left at the
current moment.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
RLIMIT_SIZE in needed only for execution from an io-wq context, hence
move all preparations from hot path to io-wq work setup.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Every call to io_req_defer_prep() is prepended with allocating ->io,
just do that in the function. And while we're at it, mark error paths
with unlikey and replace "if (ret < 0)" with "if (ret)".
There is only one change in the observable behaviour, that's instead of
killing the head request right away on error, it postpones it until the
link is assembled, that looks more preferable.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A switch in __io_clean_op() doesn't have default, it's pointless to list
opcodes that doesn't do any cleanup. Remove IORING_OP_OPEN* from there.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The only caller of io_req_work_grab_env() is io_prep_async_work(), and
they are both initialising req->work. Inline grab_env(), it's easier
to keep this way, moreover there already were bugs with misplacing
io_req_init_async().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Create the same logic flow for the write() interface as we have for the
ioctl() path by making sure that the object is committed or aborted
automatically after HW object creation.
Link: https://lore.kernel.org/r/20200719052223.75245-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
req->cflags is used only for defer-completion path, just use completion
data to store it. With the 4 bytes from the ->sequence patch and
compacting io_kiocb, this frees 8 bytes.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
req->sequence is used only for deferred (i.e. DRAIN) requests, but
initialised for every request. Remove req->sequence from io_kiocb
together with its initialisation in io_init_req().
Replace it with a new field in struct io_defer_entry, that will be
calculated only when needed in io_req_defer(), which is a slow path.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The only left user of req->list is DRAIN, hence instead of keeping a
separate per request list for it, do that with old fashion non-intrusive
lists allocated on demand. That's a really slow path, so that's OK.
This removes req->list and so sheds 16 bytes from io_kiocb.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Instead of using shared req->list, hang timeouts up on their own list
entry. struct io_timeout have enough extra space for it, but if that
will be a problem ->inflight_entry can reused for that.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As with the completion path, also use compl.list for overflowed
requests. If cleaned up properly, nobody needs per-op data there
anymore.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
req->inflight_entry is used to track requests that grabbed files_struct.
Let's share it with iopoll list, because the only iopoll'ed ops are
reads and writes, which don't need a file table.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It supports both polling and I/O polling. Rename ctx->poll to clearly
show that it's only in I/O poll case.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Calling io_req_complete(req) means that the request is done, and there
is nothing left but to clean it up. That also means that per-op data
after that should not be used, so we're free to reuse it in completion
path, e.g. to store overflow_list as done in this patch.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As for import_iovec(), return !=NULL iovec from io_import_iovec() only
when it should be freed. That includes returning NULL when iovec is
already in req->io, because it should be deallocated by other means,
e.g. inside op handler. After io_setup_async_rw() local iovec to ->io,
just mark it NULL, to follow the idea in io_{read,write} as well.
That's easier to follow, and especially useful if we want to reuse
per-op space for completion data.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
[axboe: only call kfree() on non-NULL pointer]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Preparing reads/writes for async is a bit tricky. Extract a helper to
not repeat it twice.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Don't deref req->io->rw every time, but put it in a local variable. This
looks prettier, generates less instructions, and doesn't break alias
analysis.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_kiocb::task_work was de-unionised, and is not planned to be shared
back, because it's too useful and commonly used. Hence, instead of
keeping a separate task_work in struct io_async_rw just reuse
req->task_work.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
send/recv msghdr initialisation works with struct io_async_msghdr, but
pulls the whole struct io_async_ctx for no reason. That complicates it
with composite accessing, e.g. io->msg.
Use and pass the most specific type, which is struct io_async_msghdr.
It is the larget field in union io_async_ctx and doesn't save stack
space, but looks clearer.
The most of the changes are replacing "io->msg." with "iomsg->"
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Every second field in send/recv is called msg, make it a bit more
understandable by renaming ->msg, which is a user provided ptr,
to ->umsg.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
rings_size() sets sq_offset to the total size of the rings (the returned
value which is used for memory allocation). This is wrong: sq array should
be located within the rings, not after them. Set sq_offset to where it
should be.
Fixes: 75b28affdd ("io_uring: allocate the two rings together")
Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
Acked-by: Hristo Venev <hristo@venev.name>
Cc: io-uring@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The devlink device name is of the form "supplier:consumer". But ":" is
fairly common in device names and makes it visually hard to distinguish
supplier and consumer. So, replace it with "--" to make it easier.
Signed-off-by: Saravana Kannan <saravanak@google.com>
Link: https://lore.kernel.org/r/20200724180523.1393383-1-saravanak@google.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Merge in io_uring-5.8 fixes, as changes/cleanups to how we do locked
mem accounting require a fixup, and only one of the spots are noticed
by git as the other merges cleanly. The flags fix from io_uring-5.8
also causes a merge conflict, the leak fix for recvmsg, the double poll
fix, and the link failure locking fix.
* io_uring-5.8:
io_uring: fix lockup in io_fail_links()
io_uring: fix ->work corruption with poll_add
io_uring: missed req_init_async() for IOSQE_ASYNC
io_uring: always allow drain/link/hardlink/async sqe flags
io_uring: ensure double poll additions work with both request types
io_uring: fix recvmsg memory leak with buffer selection
io_uring: fix not initialised work->flags
io_uring: fix missing msg_name assignment
io_uring: account user memory freed when exit has been queued
io_uring: fix memleak in io_sqe_files_register()
io_uring: fix memleak in __io_sqe_files_update()
io_uring: export cq overflow status to userspace
Signed-off-by: Jens Axboe <axboe@kernel.dk>
req->work might be already initialised by the time it gets into
__io_arm_poll_handler(), which will corrupt it by using fields that are
in an union with req->work. Luckily, the only side effect is missing
put_creds(). Clean req->work before going there.
Suggested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Muchun Song