Commit fb5ccc98 ("io_uring: Fix broken links with offloading")
introduced a potential performance regression with unconditionally
taking mm even for READ/WRITE_FIXED operations.

Return the logic handling it back. mm-faulted requests will go through
the generic submission path, so honoring links and drains, but will
fail further on req->has_user check.

Fixes: fb5ccc98 ("io_uring: Fix broken links with offloading")
Cc: stable@vger.kernel.org # v5.4
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

submit->index is used only for inbound check in submission path (i.e.
head < ctx->sq_entries). However, it always will be true, as
1. it's already validated by io_get_sqring()
2. ctx->sq_entries can't be changedd in between, because of held
ctx->uring_lock and ctx->refs.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

To trace io_uring activity one can get an information from workqueue and
io trace events, but looks like some parts could be hard to identify via
this approach. Making what happens inside io_uring more transparent is
important to be able to reason about many aspects of it, hence introduce
the set of tracing events.

All such events could be roughly divided into two categories:

* those, that are helping to understand correctness (from both kernel
  and an application point of view). E.g. a ring creation, file
  registration, or waiting for available CQE. Proposed approach is to
  get a pointer to an original structure of interest (ring context, or
  request), and then find relevant events. io_uring_queue_async_work
  also exposes a pointer to work_struct, to be able to track down
  corresponding workqueue events.

* those, that provide performance related information. Mostly it's about
  events that change the flow of requests, e.g. whether an async work
  was queued, or delayed due to some dependencies. Another important
  case is how io_uring optimizations (e.g. registered files) are
  utilized.
Signed-off-by: NDmitrii Dolgov <9erthalion6@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We might have cases where the need for a specific timeout is gone, add
support for canceling an existing timeout operation. This works like the
POLL_REMOVE command, where the application passes in the user_data of
the timeout it wishes to cancel in the sqe->addr field.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This is a pretty trivial addition on top of the relative timeouts
we have now, but it's handy for ensuring tighter timing for those
that are building scheduling primitives on top of io_uring.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

There is no function change, just to clean up the code, use s->in_async
to make the code know where it is.
Signed-off-by: NJackie Liu <liuyun01@kylinos.cn>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We currently size the CQ ring as twice the SQ ring, to allow some
flexibility in not overflowing the CQ ring. This is done because the
SQE life time is different than that of the IO request itself, the SQE
is consumed as soon as the kernel has seen the entry.

Certain application don't need a huge SQ ring size, since they just
submit IO in batches. But they may have a lot of requests pending, and
hence need a big CQ ring to hold them all. By allowing the application
to control the CQ ring size multiplier, we can cater to those
applications more efficiently.

If an application wants to define its own CQ ring size, it must set
IORING_SETUP_CQSIZE in the setup flags, and fill out
io_uring_params->cq_entries. The value must be a power of two.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Allows the application to remove/replace/add files to/from a file set.
Passes in a struct:

struct io_uring_files_update {
	__u32 offset;
	__s32 *fds;
};

that holds an array of fds, size of array passed in through the usual
nr_args part of the io_uring_register() system call. The logic is as
follows:

1) If ->fds[i] is -1, the existing file at i + ->offset is removed from
   the set.
2) If ->fds[i] is a valid fd, the existing file at i + ->offset is
   replaced with ->fds[i].

For case #2, is the existing file is currently empty (fd == -1), the
new fd is simply added to the array.
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This is in preparation for allowing updates to fixed file sets without
requiring a full unregister+register.
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Currently any dependent link is executed from a new workqueue context,
which means that we'll be doing a context switch per link in the chain.
If we are running the completion of the current request from our async
workqueue and find that the next request is a link, then run it directly
from the workqueue context instead of forcing another switch.

This improves the performance of linked SQEs, and reduces the CPU
overhead.
Reviewed-by: NJackie Liu <liuyun01@kylinos.cn>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

syzkaller reported an issue where it looks like a malicious app can
trigger a use-after-free of reading the ctx ->sq_array and ->rings
value right after having installed the ring fd in the process file
table.

Defer ring fd installation until after we're done reading those
values.

Fixes: 75b28aff ("io_uring: allocate the two rings together")
Reported-by: syzbot+6f03d895a6cd0d06187f@syzkaller.appspotmail.com
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_queue_link_head() owns shadow_req after taking it as an argument.
By not freeing it in case of an error, it can leak the request along
with taken ctx->refs.
Reviewed-by: NJackie Liu <liuyun01@kylinos.cn>
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We currently assume that submissions from the sqthread are successful,
and if IO polling is enabled, we use that value for knowing how many
completions to look for. But if we overflowed the CQ ring or some
requests simply got errored and already completed, they won't be
available for polling.

For the case of IO polling and SQTHREAD usage, look at the pending
poll list. If it ever hits empty then we know that we don't have
anymore pollable requests inflight. For that case, simply reset
the inflight count to zero.
Reported-by: NPavel Begunkov <asml.silence@gmail.com>
Reviewed-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We currently use the ring values directly, but that can lead to issues
if the application is malicious and changes these values on our behalf.
Created in-kernel cached versions of them, and just overwrite the user
side when we update them. This is similar to how we treat the sq/cq
ring tail/head updates.
Reported-by: NPavel Begunkov <asml.silence@gmail.com>
Reviewed-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_ring_submit() finalises with
1. io_commit_sqring(), which releases sqes to the userspace
2. Then calls to io_queue_link_head(), accessing released head's sqe

Reorder them.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_sq_thread() processes sqes by 8 without considering links. As a
result, links will be randomely subdivided.

The easiest way to fix it is to call io_get_sqring() inside
io_submit_sqes() as do io_ring_submit().

Downsides:
1. This removes optimisation of not grabbing mm_struct for fixed files
2. It submitting all sqes in one go, without finer-grained sheduling
with cq processing.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

There is a bug, where failed linked requests are returned not with
specified @user_data, but with garbage from a kernel stack.

The reason is that io_fail_links() uses req->user_data, which is
uninitialised when called from io_queue_sqe() on fail path.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

The sequence number of the timeout req (req->sequence) indicate the
expected completion request. Because of each timeout req consume a
sequence number, so the sequence of each timeout req on the timeout
list shouldn't be the same. But now, we may get the same number (also
incorrect) if we insert a new entry before the last one, such as submit
such two timeout reqs on a new ring instance below.

                    req->sequence
 req_1 (count = 2):       2
 req_2 (count = 1):       2

Then, if we submit a nop req, req_2 will still timeout even the nop req
finished. This patch fix this problem by adjust the sequence number of
each reordered reqs when inserting a new entry.
Signed-off-by: Nzhangyi (F) <yi.zhang@huawei.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

The sequence number of reqs on the timeout_list before the timeout req
should be adjusted in io_timeout_fn(), because the current timeout req
will consumes a slot in the cq_ring and cq_tail pointer will be
increased, otherwise other timeout reqs may return in advance without
waiting for enough wait_nr.
Signed-off-by: Nzhangyi (F) <yi.zhang@huawei.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

There are cases where it isn't always safe to block for submission,
even if the caller asked to wait for events as well. Revert the
previous optimization of doing that.

This reverts two commits:

bf7ec93c
c5766668

Fixes: c5766668 ("io_uring: optimize submit_and_wait API")
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If ctx->cached_sq_head < nxt_sq_head, we should add UINT_MAX to tmp, not
tmp_nxt.

Fixes: 5da0fb1a ("io_uring: consider the overflow of sequence for timeout req")
Signed-off-by: Nyangerkun <yangerkun@huawei.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We've got two issues with the non-regular file handling for non-blocking
IO:

1) We don't want to re-do a short read in full for a non-regular file,
   as we can't just read the data again.
2) For non-regular files that don't support non-blocking IO attempts,
   we need to punt to async context even if the file is opened as
   non-blocking. Otherwise the caller always gets -EAGAIN.

Add two new request flags to handle these cases. One is just a cache
of the inode S_ISREG() status, the other tells io_uring that we always
need to punt this request to async context, even if REQ_F_NOWAIT is set.

Cc: stable@vger.kernel.org
Reported-by: NHrvoje Zeba <zeba.hrvoje@gmail.com>
Tested-by: NHrvoje Zeba <zeba.hrvoje@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Now we recalculate the sequence of timeout with 'req->sequence =
ctx->cached_sq_head + count - 1', judge the right place to insert
for timeout_list by compare the number of request we still expected for
completion. But we have not consider about the situation of overflow:

1. ctx->cached_sq_head + count - 1 may overflow. And a bigger count for
the new timeout req can have a small req->sequence.

2. cached_sq_head of now may overflow compare with before req. And it
will lead the timeout req with small req->sequence.

This overflow will lead to the misorder of timeout_list, which can lead
to the wrong order of the completion of timeout_list. Fix it by reuse
req->submit.sequence to store the count, and change the logic of
inserting sort in io_timeout.
Signed-off-by: Nyangerkun <yangerkun@huawei.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We have two ways a request can be deferred:

1) It's a regular request that depends on another one
2) It's a timeout that tracks completions

We have a shared helper to determine whether to defer, and that
attempts to make the right decision based on the request. But we
only have some of this information in the caller. Un-share the
two timeout/defer helpers so the caller can use the right one.

Fixes: 5262f567 ("io_uring: IORING_OP_TIMEOUT support")
Reported-by: Nyangerkun <yangerkun@huawei.com>
Reviewed-by: NJackie Liu <liuyun01@kylinos.cn>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We should not remove the workqueue, we just need to ensure that the
workqueues are synced. The workqueues are torn down on ctx removal.

Cc: stable@vger.kernel.org
Fixes: 6b06314c ("io_uring: add file set registration")
Reported-by: NStefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Any changes interesting to tasks waiting in io_cqring_wait() are
commited with io_cqring_ev_posted(). However, io_ring_drop_ctx_refs()
also tries to do that but with no reason, that means spurious wakeups
every io_free_req() and io_uring_enter().

Just use percpu_ref_put() instead.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_queue_link_head() accepts @force_nonblock flag, but io_ring_submit()
passes something opposite.

Fixes: c5766668 ("io_uring: optimize submit_and_wait API")
Reported-by: Nkbuild test robot <lkp@intel.com>
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

All system calls use struct __kernel_timespec instead of the old struct
timespec, but this one was just added with the old-style ABI. Change it
now to enforce the use of __kernel_timespec, avoiding ABI confusion and
the need for compat handlers on 32-bit architectures.

Any user space caller will have to use __kernel_timespec now, but this
is unambiguous and works for any C library regardless of the time_t
definition. A nicer way to specify the timeout would have been a less
ambiguous 64-bit nanosecond value, but I suppose it's too late now to
change that as this would impact both 32-bit and 64-bit users.

Fixes: 5262f567 ("io_uring: IORING_OP_TIMEOUT support")
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

For batched IO, it's not uncommon for waiters to ask for more than 1
IO to complete before being woken up. This is a problem with
wait_event() since tasks will get woken for every IO that completes,
re-check condition, then go back to sleep. For batch counts on the
order of what you do for high IOPS, that can result in 10s of extra
wakeups for the waiting task.

Add a private wake function that checks for the wake up count criteria
being met before calling autoremove_wake_function(). Pavel reports that
one test case he has runs 40% faster with proper batching of wakeups.
Reported-by: NPavel Begunkov <asml.silence@gmail.com>
Tested-by: NPavel Begunkov <asml.silence@gmail.com>
Reviewed-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Patch series "Make working with compound pages easier", v2.

These three patches add three helpers and convert the appropriate
places to use them.

This patch (of 3):

It's unnecessarily hard to find out the size of a potentially huge page.
Replace 'PAGE_SIZE << compound_order(page)' with page_size(page).

Link: http://lkml.kernel.org/r/20190721104612.19120-2-willy@infradead.orgSigned-off-by: NMatthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: NMichal Hocko <mhocko@suse.com>
Reviewed-by: NAndrew Morton <akpm@linux-foundation.org>
Reviewed-by: NIra Weiny <ira.weiny@intel.com>
Acked-by: NKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

After 75b28aff("io_uring: allocate the two rings together"), we compare
sq.head with cached_cq_tail to determine does there any cq invalid.
Actually, we should use cq.head.

Fixes: 75b28aff ("io_uring: allocate the two rings together")
Signed-off-by: Nyangerkun <yangerkun@huawei.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Currently we just -EINVAL a read or write to an fd that isn't backed
by ->read_iter() or ->write_iter(). But we can handle them just fine,
as long as we punt fo async context first.

Implement a simple loop function for doing ->read() or ->write()
instead, and ensure we call it appropriately.
Reported-by: N李通洲 <carter.li@eoitek.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

There's been a few requests for functionality similar to io_getevents()
and epoll_wait(), where the user can specify a timeout for waiting on
events. I deliberately did not add support for this through the system
call initially to avoid overloading the args, but I can see that the use
cases for this are valid.

This adds support for IORING_OP_TIMEOUT. If a user wants to get woken
when waiting for events, simply submit one of these timeout commands
with your wait call (or before). This ensures that the application
sleeping on the CQ ring waiting for events will get woken. The timeout
command is passed in as a pointer to a struct timespec. Timeouts are
relative. The timeout command also includes a way to auto-cancel after
N events has passed.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If preempt isn't enabled in the kernel, we can run into hang issues with
sqthread submissions. Use cond_resched() to play nice instead of
cpu_relax(), if we end up starting the loop and not having any events
pending for submissions.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Sometimes io_get_req will return a NUL, then we need to do the
correct error handling, otherwise it will cause the kernel null
pointer exception.

Fixes: 4fe2c963 ("io_uring: add support for link with drain")
Signed-off-by: NJackie Liu <liuyun01@kylinos.cn>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If we end up getting woken in poll (due to a signal), then we may need
to punt the poll request to an async worker. When we do that, we look up
the list to queue at, deferefencing req->submit.sqe, however that is
only set for requests we initially decided to queue async.

This fixes a crash with poll command usage and wakeups that need to punt
to async context.

Fixes: 54a91f3b ("io_uring: limit parallelism of buffered writes")
Signed-off-by: NJens Axboe <axboe@kernel.dk>

There is a potential dangling pointer problem. we never clean
shadow_req, if there are multiple link lists in this series of
sqes, then the shadow_req will not reallocate, and continue to
use the last one. but in the previous, his memory has been
released, thus forming a dangling pointer. let's clean up him
and make sure that every new link list can reapply for a new
shadow_req.

Fixes: 4fe2c963 ("io_uring: add support for link with drain")
Signed-off-by: NJackie Liu <liuyun01@kylinos.cn>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Just clean up the code, no function changes.
Signed-off-by: NJackie Liu <liuyun01@kylinos.cn>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Some workloads can require far more than 4K oustanding entries. For
example memcached can have ~300K sockets over ~40 cores. Bumping the max
to 32K seems to work pretty well.
Reported-by: NDan Melnic <dmm@fb.com>
Signed-off-by: NDaniel Xu <dxu@dxuuu.xyz>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

The way the logic is setup in io_uring_enter() means that you can't wake
up the SQ poller thread while at the same time waiting (or polling) for
completions afterwards. There's no reason for that to be the case.
Reported-by: NLewis Baker <lbaker@fb.com>
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>