Now we've got a more generic region registration API, place
IORING_ENTER_EXT_ARG_REG and re-enable it.
First, the user has to register a region with the
IORING_MEM_REGION_REG_WAIT_ARG flag set. It can only be done for a
ring in a disabled state, aka IORING_SETUP_R_DISABLED, to avoid races
with already running waiters. With that we should have stable constant
values for ctx->cq_wait_{size,arg} in io_get_ext_arg_reg() and hence no
READ_ONCE required.
The other API difference is that we're now passing byte offsets instead
of indexes. The user _must_ align all offsets / pointers to the native
word size, failing to do so might but not necessarily has to lead to a
failure usually returned as -EFAULT. liburing will be hiding this
details from users.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/81822c1b4ffbe8ad391b4f9ad1564def0d26d990.1731689588.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Regions will serve multiple purposes. First, with it we can decouple
ring/etc. object creation from registration / mapping of the memory they
will be placed in. We already have hacks that allow to put both SQ and
CQ into the same huge page, in the future we should be able to:
region = create_region(io_ring);
create_pbuf_ring(io_uring, region, offset=0);
create_pbuf_ring(io_uring, region, offset=N);
The second use case is efficiently passing parameters. The following
patch enables back on top of regions IORING_ENTER_EXT_ARG_REG, which
optimises wait arguments. It'll also be useful for request arguments
replacing iovecs, msghdr, etc. pointers. Eventually it would also be
handy for BPF as well if it comes to fruition.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/0798cf3a14fad19cfc96fc9feca5f3e11481691d.1731689588.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We've got a good number of mappings we share with the userspace, that
includes the main rings, provided buffer rings, upcoming rings for
zerocopy rx and more. All of them duplicate user argument parsing and
some internal details as well (page pinnning, huge page optimisations,
mmap'ing, etc.)
Introduce a notion of regions. For userspace for now it's just a new
structure called struct io_uring_region_desc which is supposed to
parameterise all such mapping / queue creations. A region either
represents a user provided chunk of memory, in which case the user_addr
field should point to it, or a request for the kernel to allocate the
memory, in which case the user would need to mmap it after using the
offset returned in the mmap_offset field. With a uniform userspace API
we can avoid additional boiler plate code and apply future optimisation
to all of them at once.
Internally, there is a new structure struct io_mapped_region holding all
relevant runtime information and some helpers to work with it. This
patch limits it to user provided regions.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/0e6fe25818dfbaebd1bd90b870a6cac503fe1a24.1731689588.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We're a bit too frivolous with types of nr_pages arguments, converting
it to long and back to int, passing an unsigned int pointer as an int
pointer and so on. Shouldn't cause any problem but should be carefully
reviewed, but until then let's add a WARN_ON_ONCE check to be more
confident callers don't pass poorely checked arguents.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d48e0c097cbd90fb47acaddb6c247596510d8cfc.1731689588.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Use CLASS(fd) to get the file for sync message ring requests, rather
than open-code the file retrieval dance.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Link: https://lore.kernel.org/r/20241115034902.GP3387508@ZenIV
[axboe: make a more coherent commit message]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The following pattern becomes more and more:
+ io_req_assign_rsrc_node(&req->buf_node, node);
+ req->flags |= REQ_F_BUF_NODE;
so make it a helper, which is less fragile to use than above code, for
example, the BUF_NODE flag is even missed in current io_uring_cmd_prep().
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20241107110149.890530-4-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
`io_rsrc_node` instance won't be shared among different io_uring ctxs,
and its allocation 'ctx' is always same with the user's 'ctx', so it is
safe to pass user 'ctx' reference to rsrc helpers. Even in io_clone_buffers(),
`io_rsrc_node` instance is allocated actually for destination io_uring_ctx.
Then io_rsrc_node_ctx() can be removed, and the 8 bytes `ctx` pointer will be
removed from `io_rsrc_node` in the following patch.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20241107110149.890530-2-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When a DEFER_TASKRUN io_uring is terminating it requeues deferred task
work items as normal tw, which can further fallback to kthread
execution. Avoid this extra step and always push them to the fallback
kthread.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d1cd472cec2230c66bd1c8d412a5833f0af75384.1730772720.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
__io_napi_do_busy_loop now requires to have loop_end in its parameters.
This makes the code cleaner and also has the benefit of removing a
branch since the only caller not passing NULL for loop_end_arg is also
setting the value conditionally.
Signed-off-by: Olivier Langlois <olivier@trillion01.com>
Link: https://lore.kernel.org/r/d5b9bb91b1a08fff50525e1c18d7b4709b9ca100.1728828877.git.olivier@trillion01.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
1. move the sock->sk pointer validity test outside the function to
avoid the function call overhead and to make the function more
more reusable
2. change its name to __io_napi_add_id to be more precise about it is
doing
3. return an error code to report errors
Signed-off-by: Olivier Langlois <olivier@trillion01.com>
Link: https://lore.kernel.org/r/d637fa3b437d753c0f4e44ff6a7b5bf2c2611270.1728828877.git.olivier@trillion01.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The SQ index array consists of user provided indexes, which io_uring
then uses to index the SQ, and so it's susceptible to speculation. For
all other queues io_uring tracks heads and tails in kernel, and they
shouldn't need any special care.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/c6c7a25962924a55869e317e4fdb682dfdc6b279.1730687889.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than store the task_struct itself in struct io_kiocb, store
the io_uring specific task_struct. The life times are the same in terms
of io_uring, and this avoids doing some dereferences through the
task_struct. For the hot path of putting local task references, we can
deref req->tctx instead, which we'll need anyway in that function
regardless of whether it's local or remote references.
This is mostly straight forward, except the original task PF_EXITING
check needs a bit of tweaking. task_work is _always_ run from the
originating task, except in the fallback case, where it's run from a
kernel thread. Replace the potentially racy (in case of fallback work)
checks for req->task->flags with current->flags. It's either the still
the original task, in which case PF_EXITING will be sane, or it has
PF_KTHREAD set, in which case it's fallback work. Both cases should
prevent moving forward with the given request.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Right now the task_struct pointer is used as the key to match a task,
but in preparation for some io_kiocb changes, move it to using struct
io_uring_task instead. No functional changes intended in this patch.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently the io_rsrc_node assignment in io_kiocb is an array of two
pointers, as two nodes may be assigned to a request - one file node,
and one buffer node. However, the buffer node can co-exist with the
provided buffers, as currently it's not supported to use both provided
and registered buffers at the same time.
This crucially brings struct io_kiocb down to 4 cache lines again, as
before it spilled into the 5th cacheline.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than keep the type field separate rom ctx, use the fact that we
can encode up to 4 types of nodes in the LSB of the ctx pointer. Doesn't
reclaim any space right now on 64-bit archs, but it leaves a full int
for future use.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A new hybrid poll is implemented on the io_uring layer. Once an IO is
issued, it will not poll immediately, but rather block first and re-run
before IO complete, then poll to reap IO. While this poll method could
be a suboptimal solution when running on a single thread, it offers
performance lower than regular polling but higher than IRQ, and CPU
utilization is also lower than polling.
To use hybrid polling, the ring must be setup with both the
IORING_SETUP_IOPOLL and IORING_SETUP_HYBRID)IOPOLL flags set. Hybrid
polling has the same restrictions as IOPOLL, in that commands must
explicitly support it.
Signed-off-by: hexue <xue01.he@samsung.com>
Link: https://lore.kernel.org/r/20241101091957.564220-2-xue01.he@samsung.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently cloning a buffer table will fail if the destination already has
a table. But it should be possible to use it to replace existing elements.
Add a IORING_REGISTER_DST_REPLACE cloning flag, which if set, will allow
the destination to already having a buffer table. If that is the case,
then entries designated by offset + nr buffers will be replaced if they
already exist.
Note that it's allowed to use IORING_REGISTER_DST_REPLACE and not have
an existing table, in which case it'll work just like not having the
flag set and an empty table - it'll just assign the newly created table
for that case.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Right now buffer cloning is an all-or-nothing kind of thing - either the
whole table is cloned from a source to a destination ring, or nothing at
all.
However, it's not always desired to clone the whole thing. Allow for
the application to specify a source and destination offset, and a
number of buffers to clone. If the destination offset is non-zero, then
allocate sparse nodes upfront.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The empty node was used as a placeholder for a sparse entry, but it
didn't really solve any issues. The caller still has to check for
whether it's the empty node or not, it may as well just check for a NULL
return instead.
The dummy_ubuf was used for a sparse buffer entry, but NULL will serve
the same purpose there of ensuring an -EFAULT on attempted import.
Just use NULL for a sparse node, regardless of whether or not it's a
file or buffer resource.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Puts and reset an existing node in a slot, if one exists. Returns true
if a node was there, false if not. This helps cleanup some of the code
that does a lookup just to clear an existing node.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are lots of spots open-coding this functionality, add a generic
helper that does the node lookup in a speculation safe way.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For files, there's nr_user_files/file_table/file_data, and buffers have
nr_user_bufs/user_bufs/buf_data. There's no reason why file_table and
file_data can't be the same thing, and ditto for the buffer side. That
gets rid of more io_ring_ctx state that's in two spots rather than just
being in one spot, as it should be. Put all the registered file data in
one locations, and ditto on the buffer front.
This also avoids having both io_rsrc_data->nodes being an allocated
array, and ->user_bufs[] or ->file_table.nodes. There's no reason to
have this information duplicated. Keep it in one spot, io_rsrc_data,
along with how many resources are available.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add the empty node initializing to the preinit part of the io_kiocb
allocation, and reset them if they have been used.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than allocate an io_rsrc_node for an empty/sparse buffer entry,
add a const entry that can be used for that. This just needs checking
for writing the tag, and the put check needs to check for that sparse
node rather than NULL for validity.
This avoids allocating rsrc nodes for sparse buffer entries.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Work in progress, but get rid of the per-ring serialization of resource
nodes, like registered buffers and files. Main issue here is that one
node can otherwise hold up a bunch of other nodes from getting freed,
which is especially a problem for file resource nodes and networked
workloads where some descriptors may not see activity in a long time.
As an example, instantiate an io_uring ring fd and create a sparse
registered file table. Even 2 will do. Then create a socket and register
it as fixed file 0, F0. The number of open files in the app is now 5,
with 0/1/2 being the usual stdin/out/err, 3 being the ring fd, and 4
being the socket. Register this socket (eg "the listener") in slot 0 of
the registered file table. Now add an operation on the socket that uses
slot 0. Finally, loop N times, where each loop creates a new socket,
registers said socket as a file, then unregisters the socket, and
finally closes the socket. This is roughly similar to what a basic
accept loop would look like.
At the end of this loop, it's not unreasonable to expect that there
would still be 5 open files. Each socket created and registered in the
loop is also unregistered and closed. But since the listener socket
registered first still has references to its resource node due to still
being active, each subsequent socket unregistration is stuck behind it
for reclaim. Hence 5 + N files are still open at that point, where N is
awaiting the final put held up by the listener socket.
Rewrite the io_rsrc_node handling to NOT rely on serialization. Struct
io_kiocb now gets explicit resource nodes assigned, with each holding a
reference to the parent node. A parent node is either of type FILE or
BUFFER, which are the two types of nodes that exist. A request can have
two nodes assigned, if it's using both registered files and buffers.
Since request issue and task_work completion is both under the ring
private lock, no atomics are needed to handle these references. It's a
simple unlocked inc/dec. As before, the registered buffer or file table
each hold a reference as well to the registered nodes. Final put of the
node will remove the node and free the underlying resource, eg unmap the
buffer or put the file.
Outside of removing the stall in resource reclaim described above, it
has the following advantages:
1) It's a lot simpler than the previous scheme, and easier to follow.
No need to specific quiesce handling anymore.
2) There are no resource node allocations in the fast path, all of that
happens at resource registration time.
3) The structs related to resource handling can all get simplified
quite a bit, like io_rsrc_node and io_rsrc_data. io_rsrc_put can
go away completely.
4) Handling of resource tags is much simpler, and doesn't require
persistent storage as it can simply get assigned up front at
registration time. Just copy them in one-by-one at registration time
and assign to the resource node.
The only real downside is that a request is now explicitly limited to
pinning 2 resources, one file and one buffer, where before just
assigning a resource node to a request would pin all of them. The upside
is that it's easier to follow now, as an individual resource is
explicitly referenced and assigned to the request.
With this in place, the above mentioned example will be using exactly 5
files at the end of the loop, not N.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It's only used from __io_req_set_rsrc_node(), and it takes both the ctx
and node itself, while never using the ctx. Just open-code the basic
refs++ in __io_req_set_rsrc_node() instead.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In preparation for not pinning the whole registered file table, open
code the second potential direct file assignment. This will be handled
by appropriate helpers in the future, for now just do it manually.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Doesn't matter right now as there's still some bytes left for it, but
let's prepare for the io_kiocb potentially growing and add a specific
freeptr offset for it.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Useful for testing performance/efficiency impact of registered files
and buffers, vs (particularly) non-registered files.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Generally applications have 1 or a few waits of waiting, yet they pass
in a struct io_uring_getevents_arg every time. This needs to get copied
and, in turn, the timeout value needs to get copied.
Rather than do this for every invocation, allow the application to
register a fixed set of wait regions that can simply be indexed when
asking the kernel to wait on events.
At ring setup time, the application can register a number of these wait
regions and initialize region/index 0 upfront:
struct io_uring_reg_wait *reg;
reg = io_uring_setup_reg_wait(ring, nr_regions, &ret);
/* set timeout and mark as set, sigmask/sigmask_sz as needed */
reg->ts.tv_sec = 0;
reg->ts.tv_nsec = 100000;
reg->flags = IORING_REG_WAIT_TS;
where nr_regions >= 1 && nr_regions <= PAGE_SIZE / sizeof(*reg). The
above initializes index 0, but 63 other regions can be initialized,
if needed. Now, instead of doing:
struct __kernel_timespec timeout = { .tv_nsec = 100000, };
io_uring_submit_and_wait_timeout(ring, &cqe, nr, &t, NULL);
to wait for events for each submit_and_wait, or just wait, operation, it
can just reference the above region at offset 0 and do:
io_uring_submit_and_wait_reg(ring, &cqe, nr, 0);
to achieve the same goal of waiting 100usec without needing to copy
both struct io_uring_getevents_arg (24b) and struct __kernel_timeout
(16b) for each invocation. Struct io_uring_reg_wait looks as follows:
struct io_uring_reg_wait {
struct __kernel_timespec ts;
__u32 min_wait_usec;
__u32 flags;
__u64 sigmask;
__u32 sigmask_sz;
__u32 pad[3];
__u64 pad2[2];
};
embedding the timeout itself in the region, rather than passing it as
a pointer as well. Note that the signal mask is still passed as a
pointer, both for compatability reasons, but also because there doesn't
seem to be a lot of high frequency waits scenarios that involve setting
and resetting the signal mask for each wait.
The application is free to modify any region before a wait call, or it
can use keep multiple regions with different settings to avoid needing to
modify the same one for wait calls. Up to a page size of regions is mapped
by default, allowing PAGE_SIZE / 64 available regions for use.
The registered region must fit within a page. On a 4kb page size system,
that allows for 64 wait regions if a full page is used, as the size of
struct io_uring_reg_wait is 64b. The region registered must be aligned
to io_uring_reg_wait in size. It's valid to register less than 64
entries.
In network performance testing with zero-copy, this reduced the time
spent waiting on the TX side from 3.12% to 0.3% and the RX side from 4.4%
to 0.3%.
Wait regions are fixed for the lifetime of the ring - once registered,
they are persistent until the ring is torn down. The regions support
minimum wait timeout as well as the regular waits.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In scenarios where a high frequency of wait events are seen, the copy
of the struct io_uring_getevents_arg is quite noticeable in the
profiles in terms of time spent. It can be seen as up to 3.5-4.5%.
Rewrite the copy-in logic, saving about 0.5% of the time.
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This avoids intermediate storage for turning a __kernel_timespec
user pointer into an on-stack struct timespec64, only then to turn it
into a ktime_t.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_sqd_handle_event() just does a mutex unlock/lock dance when it's
supposed to park, somewhat relying on full ordering with the thread
trying to park it which does a similar unlock/lock dance on sqd->lock.
However, with adaptive spinning on mutexes, this can waste an awful
lot of time. Normally this isn't very noticeable, as parking and
unparking the thread isn't a common (or fast path) occurence. However,
in testing ring resizing, it's testing exactly that, as each resize
will require the SQPOLL to safely park and unpark.
Have io_sq_thread_park() explicitly wait on sqd->park_pending being
zero before attempting to grab the sqd->lock again.
In a resize test, this brings the runtime of SQPOLL down from about
60 seconds to a few seconds, just like the !SQPOLL tests. And saves
a ton of spinning time on the mutex, on both sides.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Once a ring has been created, the size of the CQ and SQ rings are fixed.
Usually this isn't a problem on the SQ ring side, as it merely controls
the available number of requests that can be submitted in a single
system call, and there's rarely a need to change that.
For the CQ ring, it's a different story. For most efficient use of
io_uring, it's important that the CQ ring never overflows. This means
that applications must size it for the worst case scenario, which can
be wasteful.
Add IORING_REGISTER_RESIZE_RINGS, which allows an application to resize
the existing rings. It takes a struct io_uring_params argument, the same
one which is used to setup the ring initially, and resizes rings
according to the sizes given.
Certain properties are always inherited from the original ring setup,
like SQE128/CQE32 and other setup options. The implementation only
allows flag associated with how the CQ ring is sized and clamped.
Existing unconsumed SQE and CQE entries are copied as part of the
process. If either the SQ or CQ resized destination ring cannot hold the
entries already present in the source rings, then the operation is failed
with -EOVERFLOW. Any register op holds ->uring_lock, which prevents new
submissions, and the internal mapping holds the completion lock as well
across moving CQ ring state.
To prevent races between mmap and ring resizing, add a mutex that's
solely used to serialize ring resize and mmap. mmap_sem can't be used
here, as as fork'ed process may be doing mmaps on the ring as well.
The ctx->resize_lock is held across mmap operations, and the resize
will grab it before swapping out the already mapped new data.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The later mapping will actually check this too, but in terms of code
clarify, explicitly check for whether or not the rings and sqes are
valid during validation. That makes it explicit that if they are
non-NULL, they are valid and can get mapped.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
