Rather than use remap_pfn_range() for this and manually free later,
switch to using vm_insert_page() and have it Just Work.
This requires a bit of effort on the mmap lookup side, as the ctx
uring_lock isn't held, which otherwise protects buffer_lists from being
torn down, and it's not safe to grab from mmap context that would
introduce an ABBA deadlock between the mmap lock and the ctx uring_lock.
Instead, lookup the buffer_list under RCU, as the the list is RCU freed
already. Use the existing reference count to determine whether it's
possible to safely grab a reference to it (eg if it's not zero already),
and drop that reference when done with the mapping. If the mmap
reference is the last one, the buffer_list and the associated memory can
go away, since the vma insertion has references to the inserted pages at
that point.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This avoids needing to care about HIGHMEM, and it makes the buffer
indexing easier as both ring provided buffer methods are now virtually
mapped in a contigious fashion.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move it into io_uring.c where it belongs, and use it in there as well
rather than have two implementations of this.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is the last holdout which does odd page checking, convert it to
vmap just like what is done for the non-mmap path.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than use remap_pfn_range() for this and manually free later,
switch to using vm_insert_pages() and have it Just Work.
If possible, allocate a single compound page that covers the range that
is needed. If that works, then we can just use page_address() on that
page. If we fail to get a compound page, allocate single pages and use
vmap() to map them into the kernel virtual address space.
This just covers the rings/sqes, the other remaining user of the mmap
remap_pfn_range() user will be converted separately. Once that is done,
we can kill the old alloc/free code.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
An identical one exists for vm_insert_page(), add one for
vm_insert_pages() to avoid needing to check for CONFIG_MMU in code using
it.
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
While valid C, anonymous enums confuse Cython (Python to C translator),
as reported by Ritesh (YoSTEALTH) [1] . Since people rely on it when
building against liburing and we want to keep this header in sync with
the library version, let's name the existing enums in the uapi header.
[1] https://github.com/cython/cython/issues/3240
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://lore.kernel.org/r/20240328210935.25640-1-krisman@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_task_work_pending() uses wq_list_empty() on ctx->work_llist, but it's
not an io_wq_work_list, it's a struct llist_head. They both have
->first as head-of-list, and it turns out the checks are identical. But
be proper and use the right helper.
Fixes: dac6a0eae7 ("io_uring: ensure iopoll runs local task work as well")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The object list is a bit of a mess, with core and opcode files mixed in.
Re-arrange it so that we have the core bits first, and then opcode
specific files after that.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The allocator will generally return memory in order, but
__io_alloc_req_refill() then adds them to a stack and we'll extract them
in the opposite order. This obviously isn't a huge deal, but:
1) it makes debugging easier when they are in order
2) keeping them in-order is the right thing to do
3) reduces the code for adding them to the stack
Just add them in reverse to the stack.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This should be plenty, rather than the default of 128, and matches what
we have on the rsrc and futex side as well.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently lists are being used to manage this, but best practice is
usually to have these in an array instead as that it cheaper to manage.
Outside of that detail, games are also played with KASAN as the list
is inside the cached entry itself.
Finally, all users of this need a struct io_cache_entry embedded in
their struct, which is union'ized with something else in there that
isn't used across the free -> realloc cycle.
Get rid of all of that, and simply have it be an array. This will not
change the memory used, as we're just trading an 8-byte member entry
for the per-elem array size.
This reduces the overhead of the recycled allocations, and it reduces
the amount of code code needed to support recycling to about half of
what it currently is.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It's now unused, drop the code related to it. This includes the
io_issue_defs->manual alloc field.
While in there, and since ->async_size is now being used a bit more
frequently and in the issue path, move it to io_issue_defs[].
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The previous commit turned on async data for uring_cmd, and did the
basic conversion of setting everything up on the prep side. However, for
a lot of use cases, -EIOCBQUEUED will get returned on issue, as the
operation got successfully queued. For that case, a persistent SQE isn't
needed, as it's just used for issue.
Unless execution goes async immediately, defer copying the double SQE
until it's necessary.
This greatly reduces the overhead of such commands, as evidenced by
a perf diff from before and after this change:
10.60% -8.58% [kernel.vmlinux] [k] io_uring_cmd_prep
where the prep side drops from 10.60% to ~2%, which is more expected.
Performance also rises from ~113M IOPS to ~122M IOPS, bringing us back
to where it was before the async command prep.
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
Reviewed-by: Anuj Gupta <anuj20.g@samsung.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Basic conversion ensuring async_data is allocated off the prep path. Adds
a basic alloc cache as well, as passthrough IO can be quite high in rate.
Tested-by: Anuj Gupta <anuj20.g@samsung.com>
Reviewed-by: Anuj Gupta <anuj20.g@samsung.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
While doing that, get rid of io_async_connect and just use the generic
io_async_msghdr. Both of them have a struct sockaddr_storage in there,
and while io_async_msghdr is bigger, if the same type can be used then
the netmsg_cache can get reused for connect as well.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Let the io_async_rw hold on to the iovec and reuse it, rather than always
allocate and free them.
Also enables KASAN for the iovec entries, so that reuse can be detected
even while they are in the cache.
While doing so, shrink io_async_rw by getting rid of the bigger embedded
fast iovec. Since iovecs are being recycled now, shrink it from 8 to 1.
This reduces the io_async_rw size from 264 to 160 bytes, a 40% reduction.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We no longer need to gate a potential retry on whether or not the
context matches our original task, as all read/write operations have
been fully prepared upfront. This means there's never any re-import
needed, and hence we can always retry requests.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
read/write requests try to put everything on the stack, and then alloc
and copy if a retry is needed. This necessitates a bunch of nasty code
that deals with intermediate state.
Get rid of this, and have the prep side setup everything that is needed
upfront, which greatly simplifies the opcode handlers.
This includes adding an alloc cache for io_async_rw, to make it cheap
to handle.
In terms of cost, this should be basically free and transparent. For
the worst case of {READ,WRITE}_FIXED which didn't need it before,
performance is unaffected in the normal peak workload that is being
used to test that. Still runs at 122M IOPS.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now that iovec recycling is being done, the iovec is no longer being
freed in there. Hence the kmsg parameter is now useless.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Right now the io_async_msghdr is recycled to avoid the overhead of
allocating+freeing it for every request. But the iovec is not included,
hence that will be allocated and freed for each transfer regardless.
This commit enables recyling of the iovec between io_async_msghdr
recycles. This avoids alloc+free for each one if an iovec is used, and
on top of that, it extends the cache hot nature of msg to the iovec as
well.
Also enables KASAN for the iovec entries, so that reuse can be detected
even while they are in the cache.
The io_async_msghdr also shrinks from 376 -> 288 bytes, an 88 byte
saving (or ~23% smaller), as the fast_iovec entry is dropped from 8
entries to a single entry. There's no point keeping a big fast iovec
entry, if iovecs aren't being allocated and freed continually.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We now ONLY call io_msg_alloc_async() from inside prep handling, which
is always locked. No need for this helper anymore, or the check in
io_msg_alloc_async() on whether the ring is locked or not.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move the io_async_msghdr out of the issue path and into prep handling,
e it's now done unconditionally and hence does not need to be part
of the issue path. This means any usage of io_sendrecv_prep_async() and
io_sendmsg_prep_async(), and hence the forced async setup path is now
unified with the normal prep setup.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move the io_async_msghdr out of the issue path and into prep handling,
since it's now done unconditionally and hence does not need to be part
of the issue path. This reduces the footprint of the multishot fast
path of multiple invocations of ->issue() per prep, and also means that
using ->prep_async() can be dropped for recvmsg asthis is now done via
setup on the prep side.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently set this separately for async/sync entry, but let's just
move it to a generic pre-issue spot and eliminate the difference
between the two.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than use an on-stack one and then need to allocate and copy if
async execution is required, always grab one upfront. This should be
very cheap, and potentially even have cache hotness benefits for
back-to-back send/recv requests.
For any recv type of request, this is probably a good choice in general,
as it's expected that no data is available initially. For send this is
not necessarily the case, as space in the socket buffer is expected to
be available. However, getting a cached io_async_msghdr is very cheap,
and as it should be cache hot, probably the difference here is neglible,
if any.
A nice side benefit is that io_setup_async_msg can get killed
completely, which has some nasty iovec manipulation code.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
No functional changes in this patch, just in preparation for carrying
more state than what is available now, if necessary.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
No functional changes in this patch, just in preparation for carrying
more state then what is being done now, if necessary. While unifying
some of this code, add a generic send setup prep handler that they can
both use.
This gets rid of some manual msghdr and sockaddr on the stack, and makes
it look a bit more like the sendmsg/recvmsg variants. Going forward, more
can get unified on top.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In practice, we just need to recycle a few elements for (by far) most
use cases. Shrink the total size down from 512 to 128, which should be
more than plenty.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For historical reasons these were special cased, as they were the only
ones that needed cancelation. But now we handle cancelations generally,
and hence there's no need to check for these in
io_ring_ctx_wait_and_kill() when io_uring_try_cancel_requests() handles
both these and the rest as well.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Make io_req_complete_post() to push all IORING_SETUP_IOPOLL requests
to task_work, it's much cleaner and should normally happen. We couldn't
do it before because there was a possibility of looping in
complete_post() -> tw -> complete_post() -> ...
Also, unexport the function and inline __io_req_complete_post().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/ea19c032ace3e0dd96ac4d991a063b0188037014.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
task_work execution is now always locked, and we shouldn't get into
io_req_complete_post() from them. That means that complete_post() is
always called out of the original task context and we don't even need to
check current.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/24ec27f27db0d8f58c974d8118dca1d345314ddc.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_post_aux_cqe(), which is used for multishot requests, delays
completions by putting CQEs into a temporary array for the purpose
completion lock/flush batching.
DEFER_TASKRUN doesn't need any locking, so for it we can put completions
directly into the CQ and defer post completion handling with a flag.
That leaves !DEFER_TASKRUN, which is not that interesting / hot for
multishot requests, so have conditional locking with deferred flush
for them.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/b1d05a81fd27aaa2a07f9860af13059e7ad7a890.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The restriction on multishot execution context disallowing io-wq is
driven by rules of io_fill_cqe_req_aux(), it should only be called in
the master task context, either from the syscall path or in task_work.
Since task_work now always takes the ctx lock implying
IO_URING_F_COMPLETE_DEFER, we can just assume that the function is
always called with its defer argument set to true.
Kill the argument. Also rename the function for more consistency as
"fill" in CQE related functions was usually meant for raw interfaces
only copying data into the CQ without any locking, waking the user
and other accounting "post" functions take care of.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/93423d106c33116c7d06bf277f651aa68b427328.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
ctx is always locked for task_work now, so get rid of struct
io_tw_state::locked. Note I'm stopping one step before removing
io_tw_state altogether, which is not empty, because it still serves the
purpose of indicating which function is a tw callback and forcing users
not to invoke them carelessly out of a wrong context. The removal can
always be done later.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/e95e1ea116d0bfa54b656076e6a977bc221392a4.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We can run normal task_work without locking the ctx, however we try to
lock anyway and most handlers prefer or require it locked. It might have
been interesting to multi-submitter ring with high contention completing
async read/write requests via task_work, however that will still need to
go through io_req_complete_post() and potentially take the lock for
rsrc node putting or some other case.
In other words, it's hard to care about it, so alawys force the locking.
The case described would also because of various io_uring caches.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/6ae858f2ef562e6ed9f13c60978c0d48926954ba.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
kiocb_done() should care to specifically redirecting requests to io-wq.
Remove the hopping to tw to then queue an io-wq, return -EAGAIN and let
the core code io_uring handle offloading.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/413564e550fe23744a970e1783dfa566291b0e6f.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
NVMe is making up issue_flags, which is a no-no in general, and to make
matters worse, they are completely the wrong ones. For a pure polled
request, which it does check for, we're already inside the
ctx->uring_lock when the completions are run off io_do_iopoll(). Hence
the correct flag would be '0' rather than IO_URING_F_UNLOCKED.
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
!IO_URING_F_UNLOCKED does not translate to availability of the deferred
completion infra, IO_URING_F_COMPLETE_DEFER does, that what we should
pass and look for to use io_req_complete_defer() and other variants.
Luckily, it's not a real problem as two wrongs actually made it right,
at least as far as io_uring_cmd_work() goes.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/aef76d34fe9410df8ecc42a14544fd76cd9d8b9e.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring cmd converts struct io_tw_state to issue_flags and later back
to io_tw_state, it's awfully ill-fated, not to mention that intermediate
issue_flags state is not correct.
Get rid of the last conversion, drag through tw everything that came
with IO_URING_F_UNLOCKED, and replace io_req_complete_defer() with a
direct call to io_req_complete_defer(), at least for the time being.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/c53fa3df749752bd058cf6f824a90704822d6bcc.1710799188.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
