io-wq assumes that work will complete fast (and not block), so it
doesn't create a new worker when work is enqueued, if we already have
at least one worker running. This is done on the assumption that if work
is running, then it will complete fast.
Add an option to force io-wq to fork a new worker for work queued. This
is signaled by setting IO_WQ_WORK_CONCURRENT on the work item. For that
case, io-wq will create a new worker, even though workers are already
running.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Not all work can be cancelled, some of it we may need to guarantee
that it runs to completion. Allow the caller to set IO_WQ_WORK_NO_CANCEL
on work that must not be cancelled. Note that the caller work function
must also check for IO_WQ_WORK_NO_CANCEL on work that is marked
IO_WQ_WORK_CANCEL.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Reschedule the current IO worker to cut the risk that it is becoming
a cpu hog.
Signed-off-by: Hillf Danton <hdanton@sina.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit e61df66c69 ("io-wq: ensure free/busy list browsing see all
items") added a list for io workers in addition to the free and busy
lists, not only making worker walk cleaner, but leaving the busy list
unused. Let's remove it.
Signed-off-by: Hillf Danton <hdanton@sina.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
- Fix a few typos found while reading the code.
- Fix stale io_get_sqring comment referencing s->sqe, the 's' parameter
was renamed to 'req', but the comment still holds.
Signed-off-by: Brian Gianforcaro <b.gianfo@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To avoid going to sleep only to get woken shortly thereafter, spin
briefly for new work upon completion of work.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We only have one cases of using the waitqueue to wake the worker, the
rest are using wake_up_process(). Since we can save some cycles not
fiddling with the waitqueue io_wqe_worker(), switch the work activation
to task wakeup and get rid of the now unused wait_queue_head_t in
struct io_worker.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently we're using 40 bytes for the io_wq_work structure, and 16 of
those is the doubly link list node. We don't need doubly linked lists,
we always add to tail to keep things ordered, and any other use case
is list traversal with deletion. For the deletion case, we can easily
support any node deletion by keeping track of the previous entry.
This shrinks io_wq_work to 32 bytes, and subsequently io_kiock from
io_uring to 216 to 208 bytes.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are several things that can go wrong in the current code on NUMA
systems, especially if not all nodes are online all the time:
- If the identifiers of the online nodes do not form a single contiguous
block starting at zero, wq->wqes will be too small, and OOB memory
accesses will occur e.g. in the loop in io_wq_create().
- If a node comes online between the call to num_online_nodes() and the
for_each_node() loop in io_wq_create(), an OOB write will occur.
- If a node comes online between io_wq_create() and io_wq_enqueue(), a
lookup is performed for an element that doesn't exist, and an OOB read
will probably occur.
Fix it by:
- using nr_node_ids instead of num_online_nodes() for the allocation size;
nr_node_ids is calculated by setup_nr_node_ids() to be bigger than the
highest node ID that could possibly come online at some point, even if
those nodes' identifiers are not a contiguous block
- creating workers for all possible CPUs, not just all online ones
This is basically what the normal workqueue code also does, as far as I can
tell.
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
These allocations are single-element allocations, so don't use the array
allocation wrapper for them.
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently pass in 4 arguments outside of the bounded size. In
preparation for adding one more argument, let's bundle them up in
a struct to make it more readable.
No functional changes in this patch.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When we find new work to process within the work handler, we queue the
linked timeout before we have issued the new work. This can be
problematic for very short timeouts, as we have a window where the new
work isn't visible.
Allow the work handler to store a callback function for this in the work
item, and flag it with IO_WQ_WORK_CB if the caller has done so. If that
is set, then io-wq will call the callback when it has setup the new work
item.
Reported-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently have a race where if setup is really slow, we can be
calling io_wq_destroy() before we're done setting up. This will cause
the caller to get stuck waiting for the manager to set things up, but
the manager already exited.
Fix this by doing a sync setup of the manager. This also fixes the case
where if we failed creating workers, we'd also get stuck.
In practice this race window was really small, as we already wait for
the manager to start. Hence someone would have to call io_wq_destroy()
after the task has started, but before it started the first loop. The
reported test case forked tons of these, which is why it became an
issue.
Reported-by: syzbot+0f1cc17f85154f400465@syzkaller.appspotmail.com
Fixes: 771b53d033 ("io-wq: small threadpool implementation for io_uring")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since we don't iterate these lists anymore after commit:
e61df66c69 ("io-wq: ensure free/busy list browsing see all items")
we don't need to retain the nulls value we use for them. That means it's
pretty pointless to wrap the hlist_nulls_head in a structure, so get rid
of it.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have two lists for workers in io-wq, a busy and a free list. For
certain operations we want to browse all workers, and we currently do
that by browsing the two separate lists. But since these lists are RCU
protected, we can potentially miss workers if they move between the two
lists while we're browsing them.
Add a third list, all_list, that simply holds all workers. A worker is
added to that list when it starts, and removed when it exits. This makes
the worker iteration cleaner, too.
Reported-by: Paul E. McKenney <paulmck@kernel.org>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
worker->cur_work is currently protected by the lock of the wqe that the
worker belongs to. When we send a signal to a worker, we need a stable
view of ->cur_work, so we need to hold that lock. But this doesn't work
so well, since we have the opposite order potentially on queueing work.
If POLL_ADD is used with a signalfd, then io_poll_wake() is called with
the signal lock, and that sometimes needs to insert work items.
Add a specific worker lock that protects the current work item. Then we
can guarantee that the task we're sending a signal is currently
processing the exact work we think it is.
Reported-by: Paul E. McKenney <paulmck@kernel.org>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For cancellation, we need to ensure that the work item stays valid for
as long as ->cur_work is valid. Right now we can't safely dereference
the work item even under the wqe->lock, because while the ->cur_work
pointer will remain valid, the work could be completing and be freed
in parallel.
Only invoke ->get/put_work() on items we know that the caller queued
themselves. Add IO_WQ_WORK_INTERNAL for io-wq to use, which is needed
when we're queueing a flush item, for instance.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring supports request types that basically have two different
lifetimes:
1) Bounded completion time. These are requests like disk reads or writes,
which we know will finish in a finite amount of time.
2) Unbounded completion time. These are generally networked IO, where we
have no idea how long they will take to complete. Another example is
POLL commands.
This patch provides support for io-wq to handle these differently, so we
don't starve bounded requests by tying up workers for too long. By default
all work is bounded, unless otherwise specified in the work item.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We hold the wqe lock at this point (which is also annotated), so there's
no need to use the careful variant of list_empty().
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We don't know what context we'll be called in for cancel, it could very
well be with IRQs disabled already. Use the IRQ saving variants of the
locking primitives.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The callback function of call_rcu() just calls kfree(), so we can use
kfree_rcu() instead of call_rcu() + callback function.
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This adds support for IORING_OP_ASYNC_CANCEL, which will attempt to
cancel requests that have been punted to async context and are now
in-flight. This works for regular read/write requests to files, as
long as they haven't been started yet. For socket based IO (or things
like accept4(2)), we can cancel work that is already running as well.
To cancel a request, the sqe must have ->addr set to the user_data of
the request it wishes to cancel. If the request is cancelled
successfully, the original request is completed with -ECANCELED
and the cancel request is completed with a result of 0. If the
request was already running, the original may or may not complete
in error. The cancel request will complete with -EALREADY for that
case. And finally, if the request to cancel wasn't found, the cancel
request is completed with -ENOENT.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is in preparation for adding opcodes that need to add new files
in a process file table, system calls like open(2) or accept4(2).
If an opcode needs this, it must set IO_WQ_WORK_NEEDS_FILES in the work
item. If work that needs to get punted to async context have this
set, the async worker will assume the original task file table before
executing the work.
Note that opcodes that need access to the current files of an
application cannot be done through IORING_SETUP_SQPOLL.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This adds support for io-wq, a smaller and specialized thread pool
implementation. This is meant to replace workqueues for io_uring. Among
the reasons for this addition are:
- We can assign memory context smarter and more persistently if we
manage the life time of threads.
- We can drop various work-arounds we have in io_uring, like the
async_list.
- We can implement hashed work insertion, to manage concurrency of
buffered writes without needing a) an extra workqueue, or b)
needlessly making the concurrency of said workqueue very low
which hurts performance of multiple buffered file writers.
- We can implement cancel through signals, for cancelling
interruptible work like read/write (or send/recv) to/from sockets.
- We need the above cancel for being able to assign and use file tables
from a process.
- We can implement a more thorough cancel operation in general.
- We need it to move towards a syslet/threadlet model for even faster
async execution. For that we need to take ownership of the used
threads.
This list is just off the top of my head. Performance should be the
same, or better, at least that's what I've seen in my testing. io-wq
supports basic NUMA functionality, setting up a pool per node.
io-wq hooks up to the scheduler schedule in/out just like workqueue
and uses that to drive the need for more/less workers.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe