struct io_rsrc_node carries a number of resources represented by struct
io_rsrc_put. That was handy before for sync overhead ammortisation, but
all complexity is gone and nodes are simple and lightweight. Let's
allocate a separate node for each resource.
Nodes and io_rsrc_put and not much different in size, and former are
cached, so node allocation should work better. That also removes some
overhead for nested iteration in io_rsrc_node_ref_zero() /
__io_rsrc_put_work().
Another reason for the patch is that it greatly reduces complexity
by moving io_rsrc_node_switch[_start]() inside io_queue_rsrc_removal(),
so users don't have to care about it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/c7d3a45b30cc14cd93700a710dd112edc703db98.1681822823.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For io_rsrc_ref_quiesce() to progress it should execute all task_work
items, including deferred ones. However, currently nobody would wake us,
and so let's set ctx->cq_wait_nr, so io_req_local_work_add() would wake
us up.
Fixes: c0e0d6ba25 ("io_uring: add IORING_SETUP_DEFER_TASKRUN")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/f1a90d1bc5ebf096475b018fed52e54f3b89d4af.1681395792.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Replace completions with waitqueues for rsrc data quiesce, the main
wakeup condition is when data refs hit zero. Note that data refs are
only changes under ->uring_lock, so we prepare before mutex_unlock()
reacquire it after taking the lock back. This change will be needed
in the next patch.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1d0dbc74b3b4fd67c8f01819e680c5e0da252956.1681395792.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use io_rsrc_node_switch() coupled with io_rsrc_node_switch_start()
for a bunch of cases including initialising ctx->rsrc_node, i.e. by
passing NULL instead of rsrc_data. Leave it to only deal with actual
node changing.
For that, first remove it from io_uring_create() and add a function
allocating the first node. Then also remove all calls to
io_rsrc_node_switch() from files/buffers register as we already have a
node installed and it does essentially nothing.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d146fe306ff98b1a5a60c997c252534f03d423d7.1681210788.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
struct delayed_work rsrc_put_work was previously used to offload node
freeing because io_rsrc_node_ref_zero() was previously called by RCU in
the IRQ context. Now, as percpu refcounting is gone, we can do it
eagerly at the spot without pushing it to a worker.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/13fb1aac1e8d068ad8fd4a0c6d0d157ab61b90c0.1680576071.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Every io_rsrc_node keeps a list of items to put, and all entries are
kmalloc()'ed. However, it's quite often to queue up only one entry per
node, so let's add an inline entry there to avoid extra allocations.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/c482c1c652c45c85ac52e67c974bc758a50fed5f.1680576071.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have too many "rsrc" around which makes the name of struct
io_rsrc_node::rsrc_list confusing. The field is responsible for keeping
a list of files or buffers, so call it item_list and add comments
around.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/3e34d4dfc1fdbb6b520f904ee6187c2ccf680efe.1680576071.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use ->rsrc_ref_lock spinlock to protect ->rsrc_ref_list in
io_rsrc_node_ref_zero(). Now we removed pcpu refcounting, which means
io_rsrc_node_ref_zero() is not executed from the irq context as an RCU
callback anymore, and we also put it under ->uring_lock.
io_rsrc_node_switch(), which queues up nodes into the list, is also
protected by ->uring_lock, so we can safely get rid of ->rsrc_ref_lock.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/6b60af883c263551190b526a55ff2c9d5ae07141.1680576071.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, for nodes we have an atomic counter and some cached
(non-atomic) refs protected by uring_lock. Let's put all ref
manipulations under uring_lock and get rid of the atomic part.
It's free as in all cases we care about we already hold the lock.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/25b142feed7d831008257d90c8b17c0115d4fc15.1680576071.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We cache refs of the current node (i.e. ctx->rsrc_node) in
ctx->rsrc_cached_refs. We'll be moving away from atomics, so move the
cached refs in struct io_rsrc_node for now. It's a prep patch and
shouldn't change anything in practise.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/9edc3669c1d71b06c2dca78b2b2b8bb9292738b9.1680576071.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
One problem with the current rsrc infra is that often updates will
generates lots of rsrc nodes, each carry pcpu refs. That takes quite a
lot of memory, especially if there is a stall, and takes lots of CPU
cycles. Only pcpu allocations takes >50 of CPU with a naive benchmark
updating files in a loop.
Replace pcpu refs with normal refcounting. There is already a hot path
avoiding atomics / refs, but following patches will further improve it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/e9ed8a9457b331a26555ff9443afc64cdaab7247.1680576071.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When fixed files are unregistered, file_alloc_end and alloc_hint
are not cleared. This can later cause a NULL pointer dereference in
io_file_bitmap_get() if auto index selection is enabled via
IORING_FILE_INDEX_ALLOC:
[ 6.519129] BUG: kernel NULL pointer dereference, address: 0000000000000000
[...]
[ 6.541468] RIP: 0010:_find_next_zero_bit+0x1a/0x70
[...]
[ 6.560906] Call Trace:
[ 6.561322] <TASK>
[ 6.561672] io_file_bitmap_get+0x38/0x60
[ 6.562281] io_fixed_fd_install+0x63/0xb0
[ 6.562851] ? __pfx_io_socket+0x10/0x10
[ 6.563396] io_socket+0x93/0xf0
[ 6.563855] ? __pfx_io_socket+0x10/0x10
[ 6.564411] io_issue_sqe+0x5b/0x3d0
[ 6.564914] io_submit_sqes+0x1de/0x650
[ 6.565452] __do_sys_io_uring_enter+0x4fc/0xb20
[ 6.566083] ? __do_sys_io_uring_register+0x11e/0xd80
[ 6.566779] do_syscall_64+0x3c/0x90
[ 6.567247] entry_SYSCALL_64_after_hwframe+0x72/0xdc
[...]
To fix the issue, set file alloc range and alloc_hint to zero after
file tables are freed.
Cc: stable@vger.kernel.org
Fixes: 4278a0deb1 ("io_uring: defer alloc_hint update to io_file_bitmap_set()")
Signed-off-by: Savino Dicanosa <sd7.dev@pm.me>
[axboe: add explicit bitmap == NULL check as well]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
| BUG: Bad page state in process kworker/u8:0 pfn:5c001
| page:00000000bfda61c8 refcount:0 mapcount:0 mapping:0000000000000000 index:0x20001 pfn:0x5c001
| head:0000000011409842 order:9 entire_mapcount:0 nr_pages_mapped:0 pincount:1
| anon flags: 0x3fffc00000b0004(uptodate|head|mappedtodisk|swapbacked|node=0|zone=0|lastcpupid=0xffff)
| raw: 03fffc0000000000 fffffc0000700001 ffffffff00700903 0000000100000000
| raw: 0000000000000200 0000000000000000 00000000ffffffff 0000000000000000
| head: 03fffc00000b0004 dead000000000100 dead000000000122 ffff00000a809dc1
| head: 0000000000020000 0000000000000000 00000000ffffffff 0000000000000000
| page dumped because: nonzero pincount
| CPU: 3 PID: 9 Comm: kworker/u8:0 Not tainted 6.3.0-rc2-00001-gc6811bf0cd87 #1
| Hardware name: linux,dummy-virt (DT)
| Workqueue: events_unbound io_ring_exit_work
| Call trace:
| dump_backtrace+0x13c/0x208
| show_stack+0x34/0x58
| dump_stack_lvl+0x150/0x1a8
| dump_stack+0x20/0x30
| bad_page+0xec/0x238
| free_tail_pages_check+0x280/0x350
| free_pcp_prepare+0x60c/0x830
| free_unref_page+0x50/0x498
| free_compound_page+0xcc/0x100
| free_transhuge_page+0x1f0/0x2b8
| destroy_large_folio+0x80/0xc8
| __folio_put+0xc4/0xf8
| gup_put_folio+0xd0/0x250
| unpin_user_page+0xcc/0x128
| io_buffer_unmap+0xec/0x2c0
| __io_sqe_buffers_unregister+0xa4/0x1e0
| io_ring_exit_work+0x68c/0x1188
| process_one_work+0x91c/0x1a58
| worker_thread+0x48c/0xe30
| kthread+0x278/0x2f0
| ret_from_fork+0x10/0x20
Mark reports an issue with the recent patches coalescing compound pages
while registering them in io_uring. The reason is that we try to drop
excessive references with folio_put_refs(), but pages were acquired
with pin_user_pages(), which has extra accounting and so should be put
down with matching unpin_user_pages() or at least gup_put_folio().
As a fix unpin_user_pages() all but first page instead, and let's figure
out a better API after.
Fixes: 57bebf807e ("io_uring/rsrc: optimise registered huge pages")
Reported-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Tested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/10efd5507d6d1f05ea0f3c601830e08767e189bd.1678980230.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The initialization assignment of the variable ret is changed to 0, only
in 'goto fail;' Use the ret variable as the function return value.
Signed-off-by: Li zeming <zeming@nfschina.com>
Link: https://lore.kernel.org/r/20230317182538.3027-1-zeming@nfschina.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Smatch complains that:
smatch warnings:
io_uring/rsrc.c:1262 io_sqe_buffer_register() error: uninitialized symbol 'folio'.
'folio' may be used uninitialized, which can happen if we end up with a
single page mapped. Ensure that we clear folio to NULL at the top so
it's always set.
Reported-by: kernel test robot <lkp@intel.com>
Reported-by: Dan Carpenter <error27@gmail.com>
Link: https://lore.kernel.org/r/202302241432.YML1CD5C-lkp@intel.com/
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When registering huge pages, internally io_uring will split them into
many PAGE_SIZE bvec entries. That's bad for performance as drivers need
to eventually dma-map the data and will do it individually for each bvec
entry. Coalesce huge pages into one large bvec.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Iterating within the first bvec entry should be essentially free, but we
use iov_iter_advance() for that, which shows up in benchmark profiles
taking up to 0.5% of CPU. Replace it with a hand coded version.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If two or more mappings go back to back to each other they can be passed
into io_uring to be registered as a single registered buffer. That would
even work if mappings came from different sources, e.g. it's possible to
mix in this way anon pages and pages from shmem or hugetlb. That is not
a problem but it'd rather be less prone if we forbid such mixing.
Cc: <stable@vger.kernel.org>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Use the bvec_set_page helper to initialize a bvec.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20230203150634.3199647-19-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The only call sites which would not allow overflow are also call sites
which would use the io_aux_cqe as they care about ordering.
So remove this parameter from io_post_aux_cqe.
Signed-off-by: Dylan Yudaken <dylany@meta.com>
Link: https://lore.kernel.org/r/20221124093559.3780686-9-dylany@meta.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Running task work when not needed can unnecessarily delay
operations. Specifically IORING_SETUP_DEFER_TASKRUN tries to avoid running
task work until the user requests it. Therefore do not run it in
io_uring_register any more.
The one catch is that io_rsrc_ref_quiesce expects it to have run in order
to process all outstanding references, and so reorder it's loop to do this.
Signed-off-by: Dylan Yudaken <dylany@meta.com>
Link: https://lore.kernel.org/r/20221107123349.4106213-1-dylany@meta.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
THe lifetime of SCM'ed files is bound to ring_sock, which is destroyed
strictly after we're done with registered file tables. This means there
is no need for the FFS_SCM hack, which was not available on 32-bit builds
anyway.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/984226a1045adf42dc35d8bd7fb5a8bbfa472ce1.1665891182.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Instead of putting io_uring's registered files in unix_gc() we want it
to be done by io_uring itself. The trick here is to consider io_uring
registered files for cycle detection but not actually putting them down.
Because io_uring can't register other ring instances, this will remove
all refs to the ring file triggering the ->release path and clean up
with io_ring_ctx_free().
Cc: stable@vger.kernel.org
Fixes: 6b06314c47 ("io_uring: add file set registration")
Reported-and-tested-by: David Bouman <dbouman03@gmail.com>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@canonical.com>
[axboe: add kerneldoc comment to skb, fold in skb leak fix]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Allow deferring async tasks until the user calls io_uring_enter(2) with
the IORING_ENTER_GETEVENTS flag. Enable this mode with a flag at
io_uring_setup time. This functionality requires that the later
io_uring_enter will be called from the same submission task, and therefore
restrict this flag to work only when IORING_SETUP_SINGLE_ISSUER is also
set.
Being able to hand pick when tasks are run prevents the problem where
there is current work to be done, however task work runs anyway.
For example, a common workload would obtain a batch of CQEs, and process
each one. Interrupting this to additional taskwork would add latency but
not gain anything. If instead task work is deferred to just before more
CQEs are obtained then no additional latency is added.
The way this is implemented is by trying to keep task work local to a
io_ring_ctx, rather than to the submission task. This is required, as the
application will want to wake up only a single io_ring_ctx at a time to
process work, and so the lists of work have to be kept separate.
This has some other benefits like not having to check the task continually
in handle_tw_list (and potentially unlocking/locking those), and reducing
locks in the submit & process completions path.
There are networking cases where using this option can reduce request
latency by 50%. For example a contrived example using [1] where the client
sends 2k data and receives the same data back while doing some system
calls (to trigger task work) shows this reduction. The reason ends up
being that if sending responses is delayed by processing task work, then
the client side sits idle. Whereas reordering the sends first means that
the client runs it's workload in parallel with the local task work.
[1]:
Using https://github.com/DylanZA/netbench/tree/defer_run
Client:
./netbench --client_only 1 --control_port 10000 --host <host> --tx "epoll --threads 16 --per_thread 1 --size 2048 --resp 2048 --workload 1000"
Server:
./netbench --server_only 1 --control_port 10000 --rx "io_uring --defer_taskrun 0 --workload 100" --rx "io_uring --defer_taskrun 1 --workload 100"
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Link: https://lore.kernel.org/r/20220830125013.570060-5-dylany@fb.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pavel Begunkov