IORING_OP_ASYNC_CANCEL tries io-wq cancellation only for current task.
If it fails go over tctx_list and try it out for every single tctx.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
1) The first problem is io_uring_cancel_sqpoll() ->
io_uring_cancel_task_requests() basically doing park(); park(); and so
hanging.
2) Another one is more subtle, when the master task is doing cancellations,
but SQPOLL task submits in-between the end of the cancellation but
before finish() requests taking a ref to the ctx, and so eternally
locking it up.
3) Yet another is a dying SQPOLL task doing io_uring_cancel_sqpoll() and
same io_uring_cancel_sqpoll() from the owner task, they race for
tctx->wait events. And there probably more of them.
Instead do SQPOLL cancellations from within SQPOLL task context via
task_work, see io_sqpoll_cancel_sync(). With that we don't need temporal
park()/unpark() during cancellation, which is ugly, subtle and anyway
doesn't allow to do io_run_task_work() properly.
io_uring_cancel_sqpoll() is called only from SQPOLL task context and
under sqd locking, so all parking is removed from there. And so,
io_sq_thread_[un]park() and io_sq_thread_stop() are not used now by
SQPOLL task, and that spare us from some headache.
Also remove ctx->sqd_list early to avoid 2). And kill tctx->sqpoll,
which is not used anymore.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
SQPOLL thread to which we're trying to attach may be going away, it's
not nice but a more serious problem is if io_sq_offload_create() sees
sqd->thread==NULL, and tries to init it with a new thread. There are
tons of ways it can be exploited or fail.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We always do complete(&sqd->startup) almost right after sqd->thread
creation, either in the success path or in io_sq_thread_finish(). It's
specifically created not started for us to be able to set some stuff
like sqd->thread and io_uring_alloc_task_context() before following
right after wake_up_new_task().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As io_uring_cancel_files() and others let SQO to run between
io_uring_try_cancel_requests(), SQO may generate new deferred requests,
so it's safer to try to cancel them in it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We bypass IOPOLL completion polling (and reaping) for the SQPOLL thread,
but if it's the thread itself invoking cancelations, then we still need
to perform it or no one will.
Fixes: 9936c7c2bc ("io_uring: deduplicate core cancellations sequence")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Earlier kernels had SQPOLL threads that could share across anything, as
we grabbed the context we needed on a per-ring basis. This is no longer
the case, so only allow attaching directly if we're in the same thread
group. That is the common use case. For non-group tasks, just setup a
new context and thread as we would've done if sharing wasn't set. This
isn't 100% ideal in terms of CPU utilization for the forked and share
case, but hopefully that isn't much of a concern. If it is, there are
plans in motion for how to improve that. Most importantly, we want to
avoid app side regressions where sharing worked before and now doesn't.
With this patch, functionality is equivalent to previous kernels that
supported IORING_SETUP_ATTACH_WQ with SQPOLL.
Reported-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use ->ctx_new_list to notify sqo about new ctx pending, then sqo
should stop and splice it to its sqd->ctx_list, paired with
->sq_thread_comp.
The last one is broken because nobody reinitialises it, and trying to
fix it would only add more complexity and bugs. And the first isn't
really needed as is done under park(), that protects from races well.
Add ctx into sqd->ctx_list directly (under park()), it's much simpler
and allows to kill both, ctx_new_list and sq_thread_comp.
note: apparently there is no real problem at the moment, because
sq_thread_comp is used only by io_sq_thread_finish() followed by
parking, where list_del(&ctx->sqd_list) removes it well regardless
whether it's in the new or the active list.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have to set ctx->sq_thread_idle before adding a ring to an SQ task,
otherwise sqd races for seeing zero and accounting it as such.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The io-wq threads were already marked as no-freeze, but the manager was
not. On resume, we perpetually have signal_pending() being true, and
hence the manager will loop and spin 100% of the time.
Just mark the tasks created by create_io_thread() as PF_NOFREEZE by
default, and remove any knowledge of it in io-wq and io_uring.
Reported-by: Kevin Locke <kevin@kevinlocke.name>
Tested-by: Kevin Locke <kevin@kevinlocke.name>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have a tiny race where io_put_sq_data() calls io_sq_thead_stop()
and finds the thread gone, but the thread has indeed not fully
exited or called complete() yet. Close it up by always having
io_sq_thread_stop() wait on completion of the exit event.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Fix the following coccicheck warning:
./fs/io_uring.c:8984:5-8: Unneeded variable: "ret". Return "0" on line
8998
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Link: https://lore.kernel.org/r/1615271441-33649-1-git-send-email-yang.lee@linux.alibaba.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we hit an error path in the function, make sure that the io_kiocb is
fully initialized at that point so that freeing the request always sees
a valid state.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Calling io_queue_next() after spin_unlock in io_req_complete_post()
races with the other side extracting and reusing this request. Hand
coded parts of io_req_find_next() considering that io_disarm_next()
and io_req_task_queue() have (and safe) to be called with
completion_lock held.
It already does io_commit_cqring() and io_cqring_ev_posted(), so just
reuse it for post io_disarm_next().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/5672a62f3150ee7c55849f40c0037655c4f2840f.1615250156.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A preparation patch placing all preparations before extracting a next
request into a separate helper io_disarm_next().
Also, don't spuriously do ev_posted in a rare case where REQ_F_FAIL_LINK
is set but there are no requests linked (i.e. after cancelling a linked
timeout or setting IOSQE_IO_LINK on a last request of a submission
batch).
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/44ecff68d6b47e1c4e6b891bdde1ddc08cfc3590.1615250156.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Don't set IO_SQ_THREAD_SHOULD_STOP when io_sq_offload_create() has
failed on io_uring_alloc_task_context() but leave everything to
io_sq_thread_finish(), because currently io_sq_thread_finish()
hangs on trying to park it. That's great it stalls there, because
otherwise the following io_sq_thread_stop() would be skipped on
IO_SQ_THREAD_SHOULD_STOP check and the sqo would race for sqd with
freeing ctx.
A simple error injection gives something like this.
[ 245.463955] INFO: task sqpoll-test-hang:523 blocked for more than 122 seconds.
[ 245.463983] Call Trace:
[ 245.463990] __schedule+0x36b/0x950
[ 245.464005] schedule+0x68/0xe0
[ 245.464013] schedule_timeout+0x209/0x2a0
[ 245.464032] wait_for_completion+0x8b/0xf0
[ 245.464043] io_sq_thread_finish+0x44/0x1a0
[ 245.464049] io_uring_setup+0x9ea/0xc80
[ 245.464058] __x64_sys_io_uring_setup+0x16/0x20
[ 245.464064] do_syscall_64+0x38/0x50
[ 245.464073] entry_SYSCALL_64_after_hwframe+0x44/0xae
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
You can't call idr_remove() from within a idr_for_each() callback,
but you can call xa_erase() from an xa_for_each() loop, so switch the
entire personality_idr from the IDR to the XArray. This manifests as a
use-after-free as idr_for_each() attempts to walk the rest of the node
after removing the last entry from it.
Fixes: 071698e13a ("io_uring: allow registering credentials")
Cc: stable@vger.kernel.org # 5.6+
Reported-by: yangerkun <yangerkun@huawei.com>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
[Pavel: rebased (creds load was moved into io_init_req())]
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/7ccff36e1375f2b0ebf73d957f037b43becc0dde.1615212806.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are enough of problems with IORING_SETUP_R_DISABLED, including the
burden of checking and kicking off the SQO task all over the codebase --
for exit/cancel/etc.
Rework it, always start the thread but don't do submit unless the flag
is gone, that's much easier.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io-wq now is per-task, so cancellations now should match against
request's ctx.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We keep running into weird dependency issues between the sqd lock and
the parking state. Disentangle the SQPOLL thread from the last bits of
the kthread parking inheritance, and just replace the parking state,
and two associated locks, with a single rw mutex. The SQPOLL thread
keeps the mutex for read all the time, except if someone has marked us
needing to park. Then we drop/re-acquire and try again.
This greatly simplifies the parking state machine (by just getting rid
of it), and makes it a lot more obvious how it works - if you need to
modify the ctx list, then you simply park the thread which will grab
the lock for writing.
Fold in fix from Hillf Danton on not setting STOP on a fatal signal.
Fixes: e54945ae94 ("io_uring: SQPOLL stop error handling fixes")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This brings the behavior back in line with what 5.11 and earlier did,
and this is no longer needed with the improved handling of creds
not needing to do unshare().
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With IORING_SETUP_ATTACH_WQ we should let __io_sq_thread() use the
initial creds from each ctx.
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_ring_exit_work() have to cancel all requests, including those staying
in io-wq, however it tries only cancellation of current tctx, which is
NULL. If we've got task==NULL, use the ctx-to-tctx map to go over all
tctx/io-wq and try cancellations on them.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use system_unbound_wq to run io_ring_exit_work(), so it's hard to
monitor whether removal hang or not. Add WARN_ONCE to catch hangs.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We don't use task file notes anymore, and no need left in indexing
task->io_uring->xa by file, and replace it with ctx. It's better
design-wise, especially since we keep a dangling file, and so have to
keep an eye on not dereferencing it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With ->flush() gone we're now leaving all uring file notes until the
task dies/execs, so the ctx will not be freed until all tasks that have
ever submit a request die. It was nicer with flush but not much, we
could have locked as described ctx in many cases.
Now we guarantee that ctx outlives all tctx in a sense that
io_ring_exit_work() waits for all tctxs to drop their corresponding
enties in ->xa, and ctx won't go away until then. Hence, additional
io_uring file reference (a.k.a. task file notes) are not needed anymore.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Another preparation patch. When full quiesce is done on ctx exit, use
task_work infra to remove corresponding to the ctx io_uring->xa entries.
For that we use the back tctx map. Also use ->in_idle to prevent
removing it while we traversing ->xa on cancellation, just ignore it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For each pair tcxt-ctx create an object and chain it into ctx, so we
have a way to traverse all tctx that are using current ctx. Preparation
patch, will be used later.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rework io_uring_del_task_file(), so it accepts an index to delete, and
it's not necessarily have to be in the ->xa. Infer file from xa_erase()
to maintain a single origin of truth.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we go async with a request, grab the creds that the task currently has
assigned and make sure that the async side switches to them. This is
handled in the same way that we do for registered personalities.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
45d189c606 ("io_uring: replace force_nonblock with flags") did
something strange for io_openat() slicing all issue_flags but
IO_URING_F_NONBLOCK. Not a bug for now, but better to just forward the
flags.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have this weird true/false return from parking, and then some of the
callers decide to look at that. It can lead to unbalanced parks and
sqd locking. Have the callers check the thread status once it's parked.
We know we have the lock at that point, so it's either valid or it's NULL.
Fix race with parking on thread exit. We need to be careful here with
ordering of the sdq->lock and the IO_SQ_THREAD_SHOULD_PARK bit.
Rename sqd->completion to sqd->parked to reflect that this is the only
thing this completion event doesn.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The callback can only be armed, if we get -EIOCBQUEUED returned. It's
important that we clear the WAITQ bit for other cases, otherwise we can
queue for async retry and filemap will assume that we're armed and
return -EAGAIN instead of just blocking for the IO.
Cc: stable@vger.kernel.org # 5.9+
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It doesn't make sense to wait for more events to come in, if we can't
even flush the overflow we already have to the ring. Return -EBUSY for
that condition, just like we do for attempts to submit with overflow
pending.
Cc: stable@vger.kernel.org # 5.11
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This allows us to do task creation and setup without needing to use
completions to try and synchronize with the starting thread. Get rid of
the old io_wq_fork_thread() wrapper, and the 'wq' and 'worker' startup
completion events - we can now do setup before the task is running.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Linked timeouts are fired asynchronously (i.e. soft-irq), and use
generic cancellation paths to do its stuff, including poking into io-wq.
The problem is that it's racy to access tctx->io_wq, as
io_uring_task_cancel() and others may be happening at this exact moment.
Mark linked timeouts with REQ_F_INLIFGHT for now, making sure there are
no timeouts before io-wq destraction.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Instead of going into request internals, like checking req->file->f_op,
do match them based on REQ_F_INFLIGHT, it's set only when we want it to
be reliably cancelled.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_dismantle_req() is always followed by io_put_task(), which already do
proper in_idle wake ups, so we can skip waking the owner task in
io_dismantle_req(). The rules are simpler now, do io_put_task() shortly
after ending a request, and it will be fine.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
__io_queue_async_work() is only called from io_queue_async_work(),
inline it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Inline io_req_clean_work(), less code and easier to analyse
tctx dependencies and refs usage.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When we cancel SQPOLL, @task in io_uring_try_cancel_requests() will
differ from current. Use the right tctx from passed in @task, and don't
forget that it can be NULL when the io_uring ctx exits.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We no longer revert the iovec on -EIOCBQUEUED, see commit ab2125df92,
and this started causing issues for IOPOLL on devies that run out of
request slots. Turns out what outside of needing a revert for those, we
also had a bug where we didn't properly setup retry inside the submission
path. That could cause re-import of the iovec, if any, and that could lead
to spurious results if the application had those allocated on the stack.
Catch -EAGAIN retry and make the iovec stable for IOPOLL, just like we do
for !IOPOLL retries.
Cc: <stable@vger.kernel.org> # 5.9+
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Reported-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now, sqo_task is used only for a warning that is not interesting anymore
since sqo_dead is gone, remove all of that including ctx->sqo_task.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As SQPOLL task doesn't poke into ->sqo_task anymore, there is no need to
kill the sqo when the master task exits. Before it was necessary to
avoid races accessing sqo_task->files with removing them.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
[axboe: don't forget to enable SQPOLL before exit, if started disabled]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we create it in a disabled state because IORING_SETUP_R_DISABLED is
set on ring creation, we need to ensure that we've kicked the thread if
we're exiting before it's been explicitly disabled. Otherwise we can run
into a deadlock where exit is waiting go park the SQPOLL thread, but the
SQPOLL thread itself is waiting to get a signal to start.
That results in the below trace of both tasks hung, waiting on each other:
INFO: task syz-executor458:8401 blocked for more than 143 seconds.
Not tainted 5.11.0-next-20210226-syzkaller #0
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:syz-executor458 state:D stack:27536 pid: 8401 ppid: 8400 flags:0x00004004
Call Trace:
context_switch kernel/sched/core.c:4324 [inline]
__schedule+0x90c/0x21a0 kernel/sched/core.c:5075
schedule+0xcf/0x270 kernel/sched/core.c:5154
schedule_timeout+0x1db/0x250 kernel/time/timer.c:1868
do_wait_for_common kernel/sched/completion.c:85 [inline]
__wait_for_common kernel/sched/completion.c:106 [inline]
wait_for_common kernel/sched/completion.c:117 [inline]
wait_for_completion+0x168/0x270 kernel/sched/completion.c:138
io_sq_thread_park fs/io_uring.c:7115 [inline]
io_sq_thread_park+0xd5/0x130 fs/io_uring.c:7103
io_uring_cancel_task_requests+0x24c/0xd90 fs/io_uring.c:8745
__io_uring_files_cancel+0x110/0x230 fs/io_uring.c:8840
io_uring_files_cancel include/linux/io_uring.h:47 [inline]
do_exit+0x299/0x2a60 kernel/exit.c:780
do_group_exit+0x125/0x310 kernel/exit.c:922
__do_sys_exit_group kernel/exit.c:933 [inline]
__se_sys_exit_group kernel/exit.c:931 [inline]
__x64_sys_exit_group+0x3a/0x50 kernel/exit.c:931
do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46
entry_SYSCALL_64_after_hwframe+0x44/0xae
RIP: 0033:0x43e899
RSP: 002b:00007ffe89376d48 EFLAGS: 00000246 ORIG_RAX: 00000000000000e7
RAX: ffffffffffffffda RBX: 00000000004af2f0 RCX: 000000000043e899
RDX: 000000000000003c RSI: 00000000000000e7 RDI: 0000000000000000
RBP: 0000000000000000 R08: ffffffffffffffc0 R09: 0000000010000000
R10: 0000000000008011 R11: 0000000000000246 R12: 00000000004af2f0
R13: 0000000000000001 R14: 0000000000000000 R15: 0000000000000001
INFO: task iou-sqp-8401:8402 can't die for more than 143 seconds.
task:iou-sqp-8401 state:D stack:30272 pid: 8402 ppid: 8400 flags:0x00004004
Call Trace:
context_switch kernel/sched/core.c:4324 [inline]
__schedule+0x90c/0x21a0 kernel/sched/core.c:5075
schedule+0xcf/0x270 kernel/sched/core.c:5154
schedule_timeout+0x1db/0x250 kernel/time/timer.c:1868
do_wait_for_common kernel/sched/completion.c:85 [inline]
__wait_for_common kernel/sched/completion.c:106 [inline]
wait_for_common kernel/sched/completion.c:117 [inline]
wait_for_completion+0x168/0x270 kernel/sched/completion.c:138
io_sq_thread+0x27d/0x1ae0 fs/io_uring.c:6717
ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294
INFO: task iou-sqp-8401:8402 blocked for more than 143 seconds.
Reported-by: syzbot+fb5458330b4442f2090d@syzkaller.appspotmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_run_ctx_fallback() can use xchg() instead of cmpxchg(). It's simpler
and faster.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There is an unlikely but possible race using a freed context. That's
because req->task_work.func() can free a request, but we won't
necessarily find a completion in submit_state.comp and so all ctx refs
may be put by the time we do mutex_lock(&ctx->uring_ctx);
There are several reasons why it can miss going through
submit_state.comp: 1) req->task_work.func() didn't complete it itself,
but punted to iowq (e.g. reissue) and it got freed later, or a similar
situation with it overflowing and getting flushed by someone else, or
being submitted to IRQ completion, 2) As we don't hold the uring_lock,
someone else can do io_submit_flush_completions() and put our ref.
3) Bugs and code obscurities, e.g. failing to propagate issue_flags
properly.
One example is as follows
CPU1 | CPU2
=======================================================================
@req->task_work.func() |
-> @req overflwed, |
so submit_state.comp,nr==0 |
| flush overflows, and free @req
| ctx refs == 0, free it
ctx is dead, but we do |
lock + flush + unlock |
So take a ctx reference for each new ctx we see in __tctx_task_work(),
and do release it until we do all our flushing.
Fixes: 65453d1efb ("io_uring: enable req cache for task_work items")
Reported-by: syzbot+a157ac7c03a56397f553@syzkaller.appspotmail.com
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
[axboe: fold in my one-liner and fix ref mismatch]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
