This is done through IORING_OP_SENDMSG. There's a new sqe->msg_flags
for the flags argument, and the msghdr struct is passed in the
sqe->addr field.
We use MSG_DONTWAIT to force an inline fast path if sendmsg() doesn't
block, and punt to async execution if it would have.
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With SQE links, we can create chains of dependent SQEs. One example
would be queueing an SQE that's a read from one file descriptor, with
the linked SQE being a write to another with the same set of buffers.
An SQE link will not stall the pipeline, it'll just ensure that
dependent SQEs aren't issued before the previous link has completed.
Any error at submission or completion time will break the chain of SQEs.
For completions, this also includes short reads or writes, as the next
SQE could depend on the previous one being fully completed.
Any SQE in a chain that gets canceled due to any of the above errors,
will get an CQE fill with -ECANCELED as the error value.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We can encounter a short read when we're doing buffered reads and the
data is partially cached. Right now we just return the short read, but
that forces the application to read that CQE, then issue another SQE
to finish the read. That read will not be cached, and hence will result
in an async punt.
It's more efficient to do that async punt from within the kernel, as
that will the not need two round trips more to the kernel.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently these functions return < 0 on error, and 0 for success.
Change that so that we return < 0 on error, but number of bytes
for success.
Some callers already treat the return value that way, others need a
slight tweak.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If io_copy_iov() fails, it will break the loop and report success,
albeit partially completed operation.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The previous patch has ensured that io_cqring_events contain
smp_rmb memory barriers, Now we can use wait_event_interruptible
to keep the code simple.
Signed-off-by: Jackie Liu <liuyun01@kylinos.cn>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Whenever smp_rmb is required to use io_cqring_events,
keep smp_rmb inside the function io_cqring_events.
Signed-off-by: Jackie Liu <liuyun01@kylinos.cn>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This fixes couple of races which lead to infinite wait of park completion
with the following backtraces:
[20801.303319] Call Trace:
[20801.303321] ? __schedule+0x284/0x650
[20801.303323] schedule+0x33/0xc0
[20801.303324] schedule_timeout+0x1bc/0x210
[20801.303326] ? schedule+0x3d/0xc0
[20801.303327] ? schedule_timeout+0x1bc/0x210
[20801.303329] ? preempt_count_add+0x79/0xb0
[20801.303330] wait_for_completion+0xa5/0x120
[20801.303331] ? wake_up_q+0x70/0x70
[20801.303333] kthread_park+0x48/0x80
[20801.303335] io_finish_async+0x2c/0x70
[20801.303336] io_ring_ctx_wait_and_kill+0x95/0x180
[20801.303338] io_uring_release+0x1c/0x20
[20801.303339] __fput+0xad/0x210
[20801.303341] task_work_run+0x8f/0xb0
[20801.303342] exit_to_usermode_loop+0xa0/0xb0
[20801.303343] do_syscall_64+0xe0/0x100
[20801.303349] entry_SYSCALL_64_after_hwframe+0x44/0xa9
[20801.303380] Call Trace:
[20801.303383] ? __schedule+0x284/0x650
[20801.303384] schedule+0x33/0xc0
[20801.303386] io_sq_thread+0x38a/0x410
[20801.303388] ? __switch_to_asm+0x40/0x70
[20801.303390] ? wait_woken+0x80/0x80
[20801.303392] ? _raw_spin_lock_irqsave+0x17/0x40
[20801.303394] ? io_submit_sqes+0x120/0x120
[20801.303395] kthread+0x112/0x130
[20801.303396] ? kthread_create_on_node+0x60/0x60
[20801.303398] ret_from_fork+0x35/0x40
o kthread_park() waits for park completion, so io_sq_thread() loop
should check kthread_should_park() along with khread_should_stop(),
otherwise if kthread_park() is called before prepare_to_wait()
the following schedule() never returns:
CPU#0 CPU#1
io_sq_thread_stop(): io_sq_thread():
while(!kthread_should_stop() && !ctx->sqo_stop) {
ctx->sqo_stop = 1;
kthread_park()
prepare_to_wait();
if (kthread_should_stop() {
}
schedule(); <<< nobody checks park flag,
<<< so schedule and never return
o if the flag ctx->sqo_stop is observed by the io_sq_thread() loop
it is quite possible, that kthread_should_park() check and the
following kthread_parkme() is never called, because kthread_park()
has not been yet called, but few moments later is is called and
waits there for park completion, which never happens, because
kthread has already exited:
CPU#0 CPU#1
io_sq_thread_stop(): io_sq_thread():
ctx->sqo_stop = 1;
while(!kthread_should_stop() && !ctx->sqo_stop) {
<<< observe sqo_stop and exit the loop
}
if (kthread_should_park())
kthread_parkme(); <<< never called, since was
<<< never parked
kthread_park() <<< waits forever for park completion
In the current patch we quit the loop by only kthread_should_park()
check (kthread_park() is synchronous, so kthread_should_stop() is
never observed), and we abandon ->sqo_stop flag, since it is racy.
At the end of the io_sq_thread() we unconditionally call parmke(),
since we've exited the loop by the park flag.
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: linux-block@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We always pass in 0 for the cqe flags argument, since the support for
"this read hit page cache" hint was dropped.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The test case we have is rightfully failing with the current kernel:
io_uring_setup(1, 0x7ffe2cafebe0), flags: IORING_SETUP_SQPOLL|IORING_SETUP_SQ_AFF, resv: 0x00000000 0x00000000 0x00000000 0x00000000 0x00000000, sq_thread_cpu: 4
expected -1, got 3
This is in a vm, and CPU3 is the last valid one, hence asking for 4
should fail the setup with -EINVAL, not succeed. The problem is that
we're using array_index_nospec() with nr_cpu_ids as the index, hence we
wrap and end up using CPU0 instead of CPU4. This makes the setup
succeed where it should be failing.
We don't need to use array_index_nospec() as we're not indexing any
array with this. Instead just compare with nr_cpu_ids directly. This
is fine as we're checking with cpu_online() afterwards.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pach series "Add FOLL_LONGTERM to GUP fast and use it".
HFI1, qib, and mthca, use get_user_pages_fast() due to its performance
advantages. These pages can be held for a significant time. But
get_user_pages_fast() does not protect against mapping FS DAX pages.
Introduce FOLL_LONGTERM and use this flag in get_user_pages_fast() which
retains the performance while also adding the FS DAX checks. XDP has also
shown interest in using this functionality.[1]
In addition we change get_user_pages() to use the new FOLL_LONGTERM flag
and remove the specialized get_user_pages_longterm call.
[1] https://lkml.org/lkml/2019/3/19/939
"longterm" is a relative thing and at this point is probably a misnomer.
This is really flagging a pin which is going to be given to hardware and
can't move. I've thought of a couple of alternative names but I think we
have to settle on if we are going to use FL_LAYOUT or something else to
solve the "longterm" problem. Then I think we can change the flag to a
better name.
Secondly, it depends on how often you are registering memory. I have
spoken with some RDMA users who consider MR in the performance path...
For the overall application performance. I don't have the numbers as the
tests for HFI1 were done a long time ago. But there was a significant
advantage. Some of which is probably due to the fact that you don't have
to hold mmap_sem.
Finally, architecturally I think it would be good for everyone to use
*_fast. There are patches submitted to the RDMA list which would allow
the use of *_fast (they reworking the use of mmap_sem) and as soon as they
are accepted I'll submit a patch to convert the RDMA core as well. Also
to this point others are looking to use *_fast.
As an aside, Jasons pointed out in my previous submission that *_fast and
*_unlocked look very much the same. I agree and I think further cleanup
will be coming. But I'm focused on getting the final solution for DAX at
the moment.
This patch (of 7):
This patch starts a series which aims to support FOLL_LONGTERM in
get_user_pages_fast(). Some callers who would like to do a longterm (user
controlled pin) of pages with the fast variant of GUP for performance
purposes.
Rather than have a separate get_user_pages_longterm() call, introduce
FOLL_LONGTERM and change the longterm callers to use it.
This patch does not change any functionality. In the short term
"longterm" or user controlled pins are unsafe for Filesystems and FS DAX
in particular has been blocked. However, callers of get_user_pages_fast()
were not "protected".
FOLL_LONGTERM can _only_ be supported with get_user_pages[_fast]() as it
requires vmas to determine if DAX is in use.
NOTE: In merging with the CMA changes we opt to change the
get_user_pages() call in check_and_migrate_cma_pages() to a call of
__get_user_pages_locked() on the newly migrated pages. This makes the
code read better in that we are calling __get_user_pages_locked() on the
pages before and after a potential migration.
As a side affect some of the interfaces are cleaned up but this is not the
primary purpose of the series.
In review[1] it was asked:
<quote>
> This I don't get - if you do lock down long term mappings performance
> of the actual get_user_pages call shouldn't matter to start with.
>
> What do I miss?
A couple of points.
First "longterm" is a relative thing and at this point is probably a
misnomer. This is really flagging a pin which is going to be given to
hardware and can't move. I've thought of a couple of alternative names
but I think we have to settle on if we are going to use FL_LAYOUT or
something else to solve the "longterm" problem. Then I think we can
change the flag to a better name.
Second, It depends on how often you are registering memory. I have spoken
with some RDMA users who consider MR in the performance path... For the
overall application performance. I don't have the numbers as the tests
for HFI1 were done a long time ago. But there was a significant
advantage. Some of which is probably due to the fact that you don't have
to hold mmap_sem.
Finally, architecturally I think it would be good for everyone to use
*_fast. There are patches submitted to the RDMA list which would allow
the use of *_fast (they reworking the use of mmap_sem) and as soon as they
are accepted I'll submit a patch to convert the RDMA core as well. Also
to this point others are looking to use *_fast.
As an asside, Jasons pointed out in my previous submission that *_fast and
*_unlocked look very much the same. I agree and I think further cleanup
will be coming. But I'm focused on getting the final solution for DAX at
the moment.
</quote>
[1] https://lore.kernel.org/lkml/20190220180255.GA12020@iweiny-DESK2.sc.intel.com/T/#md6abad2569f3bf6c1f03686c8097ab6563e94965
[ira.weiny@intel.com: v3]
Link: http://lkml.kernel.org/r/20190328084422.29911-2-ira.weiny@intel.com
Link: http://lkml.kernel.org/r/20190328084422.29911-2-ira.weiny@intel.com
Link: http://lkml.kernel.org/r/20190317183438.2057-2-ira.weiny@intel.com
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Rich Felker <dalias@libc.org>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: James Hogan <jhogan@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Mike Marshall <hubcap@omnibond.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When punting to workers the SQE gets copied after the initial try.
There is a race condition between reading SQE data for the initial try
and copying it for punting it to the workers.
For example io_rw_done calls kiocb->ki_complete even if it was prepared
for IORING_OP_FSYNC (and would be NULL).
The easiest solution for now is to alway prepare again in the worker.
req->file is safe to prepare though as long as it is checked before use.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This issue is found by running liburing/test/io_uring_setup test.
When test run, the testcase "attempt to bind to invalid cpu" would not
pass with messages like:
io_uring_setup(1, 0xbfc2f7c8), \
flags: IORING_SETUP_SQPOLL|IORING_SETUP_SQ_AFF, \
resv: 0x00000000 0x00000000 0x00000000 0x00000000 0x00000000, \
sq_thread_cpu: 2
expected -1, got 3
FAIL
On my system, there is:
CPU(s) possible : 0-3
CPU(s) online : 0-1
CPU(s) offline : 2-3
CPU(s) present : 0-1
The sq_thread_cpu 2 is offline on my system, so the bind should fail.
But cpu_possible() will pass the check. We shouldn't be able to bind
to an offline cpu. Use cpu_online() to do the check.
After the change, the testcase run as expected: EINVAL will be returned
for cpu offlined.
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently variable ret is declared in a while-loop code block that
shadows another variable ret. When an error occurs in the while-loop
the error return in ret is not being set in the outer code block and
so the error check on ret is always going to be checking on the wrong
ret variable resulting in check that is always going to be true and
a premature return occurs.
Fix this by removing the declaration of the inner while-loop variable
ret so that shadowing does not occur.
Addresses-Coverity: ("'Constant' variable guards dead code")
Fixes: 6b06314c47 ("io_uring: add file set registration")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
No need to set it in io_poll_add; io_poll_complete doesn't use it to set
the result in the CQE.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Allow registration of an eventfd, which will trigger an event every
time a completion event happens for this io_uring instance.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are no ordering constraints between the submission and completion
side of io_uring. But sometimes that would be useful to have. One common
example is doing an fsync, for instance, and have it ordered with
previous writes. Without support for that, the application must do this
tracking itself.
This adds a general SQE flag, IOSQE_IO_DRAIN. If a command is marked
with this flag, then it will not be issued before previous commands have
completed, and subsequent commands submitted after the drain will not be
issued before the drain is started.. If there are no pending commands,
setting this flag will not change the behavior of the issue of the
command.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In io_sqe_buffer_register() we allocate a number of arrays based on the
iov_len from the user-provided iov. While we limit iov_len to SZ_1G,
we can still attempt to allocate arrays exceeding MAX_ORDER.
On a 64-bit system with 4KiB pages, for an iov where iov_base = 0x10 and
iov_len = SZ_1G, we'll calculate that nr_pages = 262145. When we try to
allocate a corresponding array of (16-byte) bio_vecs, requiring 4194320
bytes, which is greater than 4MiB. This results in SLUB warning that
we're trying to allocate greater than MAX_ORDER, and failing the
allocation.
Avoid this by using kvmalloc() for allocations dependent on the
user-provided iov_len. At the same time, fix a leak of imu->bvec when
registration fails.
Full splat from before this patch:
WARNING: CPU: 1 PID: 2314 at mm/page_alloc.c:4595 __alloc_pages_nodemask+0x7ac/0x2938 mm/page_alloc.c:4595
Kernel panic - not syncing: panic_on_warn set ...
CPU: 1 PID: 2314 Comm: syz-executor326 Not tainted 5.1.0-rc7-dirty #4
Hardware name: linux,dummy-virt (DT)
Call trace:
dump_backtrace+0x0/0x2f0 include/linux/compiler.h:193
show_stack+0x20/0x30 arch/arm64/kernel/traps.c:158
__dump_stack lib/dump_stack.c:77 [inline]
dump_stack+0x110/0x190 lib/dump_stack.c:113
panic+0x384/0x68c kernel/panic.c:214
__warn+0x2bc/0x2c0 kernel/panic.c:571
report_bug+0x228/0x2d8 lib/bug.c:186
bug_handler+0xa0/0x1a0 arch/arm64/kernel/traps.c:956
call_break_hook arch/arm64/kernel/debug-monitors.c:301 [inline]
brk_handler+0x1d4/0x388 arch/arm64/kernel/debug-monitors.c:316
do_debug_exception+0x1a0/0x468 arch/arm64/mm/fault.c:831
el1_dbg+0x18/0x8c
__alloc_pages_nodemask+0x7ac/0x2938 mm/page_alloc.c:4595
alloc_pages_current+0x164/0x278 mm/mempolicy.c:2132
alloc_pages include/linux/gfp.h:509 [inline]
kmalloc_order+0x20/0x50 mm/slab_common.c:1231
kmalloc_order_trace+0x30/0x2b0 mm/slab_common.c:1243
kmalloc_large include/linux/slab.h:480 [inline]
__kmalloc+0x3dc/0x4f0 mm/slub.c:3791
kmalloc_array include/linux/slab.h:670 [inline]
io_sqe_buffer_register fs/io_uring.c:2472 [inline]
__io_uring_register fs/io_uring.c:2962 [inline]
__do_sys_io_uring_register fs/io_uring.c:3008 [inline]
__se_sys_io_uring_register fs/io_uring.c:2990 [inline]
__arm64_sys_io_uring_register+0x9e0/0x1bc8 fs/io_uring.c:2990
__invoke_syscall arch/arm64/kernel/syscall.c:35 [inline]
invoke_syscall arch/arm64/kernel/syscall.c:47 [inline]
el0_svc_common.constprop.0+0x148/0x2e0 arch/arm64/kernel/syscall.c:83
el0_svc_handler+0xdc/0x100 arch/arm64/kernel/syscall.c:129
el0_svc+0x8/0xc arch/arm64/kernel/entry.S:948
SMP: stopping secondary CPUs
Dumping ftrace buffer:
(ftrace buffer empty)
Kernel Offset: disabled
CPU features: 0x002,23000438
Memory Limit: none
Rebooting in 1 seconds..
Fixes: edafccee56 ("io_uring: add support for pre-mapped user IO buffers")
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-block@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we don't end up actually calling submit in io_sq_wq_submit_work(),
we still need to drop the submit reference to the request. If we
don't, then we can leak the request. This can happen if we race
with ring shutdown while flushing the workqueue for requests that
require use of the mm_struct.
Fixes: e65ef56db4 ("io_uring: use regular request ref counts")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In io_sq_offload_start(), we call cpu_possible() on an unbounded cpu
value from userspace. On v5.1-rc7 on arm64 with
CONFIG_DEBUG_PER_CPU_MAPS, this results in a splat:
WARNING: CPU: 1 PID: 27601 at include/linux/cpumask.h:121 cpu_max_bits_warn include/linux/cpumask.h:121 [inline]
There was an attempt to fix this in commit:
917257daa0 ("io_uring: only test SQPOLL cpu after we've verified it")
... by adding a check after the cpu value had been limited to NR_CPU_IDS
using array_index_nospec(). However, this left an unbound check at the
start of the function, for which the warning still fires.
Let's fix this correctly by checking that the cpu value is bound by
nr_cpu_ids before passing it to cpu_possible(). Note that only
nr_cpu_ids of a cpumask are guaranteed to exist at runtime, and
nr_cpu_ids can be significantly smaller than NR_CPUs. For example, an
arm64 defconfig has NR_CPUS=256, while my test VM has 4 vCPUs.
Following the intent from the commit message for 917257daa0, the
check is moved under the SQ_AFF branch, which is the only branch where
the cpu values is consumed. The check is performed before bounding the
value with array_index_nospec() so that we don't silently accept bogus
cpu values from userspace, where array_index_nospec() would force these
values to 0.
I suspect we can remove the array_index_nospec() call entirely, but I've
conservatively left that in place, updated to use nr_cpu_ids to match
the prior check.
Tested on arm64 with the Syzkaller reproducer:
https://syzkaller.appspot.com/bug?extid=cd714a07c6de2bc34293https://syzkaller.appspot.com/x/repro.syz?x=15d8b397200000
Full splat from before this patch:
WARNING: CPU: 1 PID: 27601 at include/linux/cpumask.h:121 cpu_max_bits_warn include/linux/cpumask.h:121 [inline]
WARNING: CPU: 1 PID: 27601 at include/linux/cpumask.h:121 cpumask_check include/linux/cpumask.h:128 [inline]
WARNING: CPU: 1 PID: 27601 at include/linux/cpumask.h:121 cpumask_test_cpu include/linux/cpumask.h:344 [inline]
WARNING: CPU: 1 PID: 27601 at include/linux/cpumask.h:121 io_sq_offload_start fs/io_uring.c:2244 [inline]
WARNING: CPU: 1 PID: 27601 at include/linux/cpumask.h:121 io_uring_create fs/io_uring.c:2864 [inline]
WARNING: CPU: 1 PID: 27601 at include/linux/cpumask.h:121 io_uring_setup+0x1108/0x15a0 fs/io_uring.c:2916
Kernel panic - not syncing: panic_on_warn set ...
CPU: 1 PID: 27601 Comm: syz-executor.0 Not tainted 5.1.0-rc7 #3
Hardware name: linux,dummy-virt (DT)
Call trace:
dump_backtrace+0x0/0x2f0 include/linux/compiler.h:193
show_stack+0x20/0x30 arch/arm64/kernel/traps.c:158
__dump_stack lib/dump_stack.c:77 [inline]
dump_stack+0x110/0x190 lib/dump_stack.c:113
panic+0x384/0x68c kernel/panic.c:214
__warn+0x2bc/0x2c0 kernel/panic.c:571
report_bug+0x228/0x2d8 lib/bug.c:186
bug_handler+0xa0/0x1a0 arch/arm64/kernel/traps.c:956
call_break_hook arch/arm64/kernel/debug-monitors.c:301 [inline]
brk_handler+0x1d4/0x388 arch/arm64/kernel/debug-monitors.c:316
do_debug_exception+0x1a0/0x468 arch/arm64/mm/fault.c:831
el1_dbg+0x18/0x8c
cpu_max_bits_warn include/linux/cpumask.h:121 [inline]
cpumask_check include/linux/cpumask.h:128 [inline]
cpumask_test_cpu include/linux/cpumask.h:344 [inline]
io_sq_offload_start fs/io_uring.c:2244 [inline]
io_uring_create fs/io_uring.c:2864 [inline]
io_uring_setup+0x1108/0x15a0 fs/io_uring.c:2916
__do_sys_io_uring_setup fs/io_uring.c:2929 [inline]
__se_sys_io_uring_setup fs/io_uring.c:2926 [inline]
__arm64_sys_io_uring_setup+0x50/0x70 fs/io_uring.c:2926
__invoke_syscall arch/arm64/kernel/syscall.c:35 [inline]
invoke_syscall arch/arm64/kernel/syscall.c:47 [inline]
el0_svc_common.constprop.0+0x148/0x2e0 arch/arm64/kernel/syscall.c:83
el0_svc_handler+0xdc/0x100 arch/arm64/kernel/syscall.c:129
el0_svc+0x8/0xc arch/arm64/kernel/entry.S:948
SMP: stopping secondary CPUs
Dumping ftrace buffer:
(ftrace buffer empty)
Kernel Offset: disabled
CPU features: 0x002,23000438
Memory Limit: none
Rebooting in 1 seconds..
Fixes: 917257daa0 ("io_uring: only test SQPOLL cpu after we've verified it")
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: linux-block@vger.kernel.org
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Simplied the logic
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently we only post a cqe if we get an error OUTSIDE of submission.
For submission, we return the error directly through io_uring_enter().
This is a bit awkward for applications, and it makes more sense to
always post a cqe with an error, if the error happens on behalf of an
sqe.
This changes submission behavior a bit. io_uring_enter() returns -ERROR
for an error, and > 0 for number of sqes submitted. Before this change,
if you wanted to submit 8 entries and had an error on the 5th entry,
io_uring_enter() would return 4 (for number of entries successfully
submitted) and rewind the sqring. The application would then have to
peek at the sqring and figure out what was wrong with the head sqe, and
then skip it itself. With this change, we'll return 5 since we did
consume 5 sqes, and the last sqe (with the error) will result in a cqe
being posted with the error.
This makes the logic easier to handle in the application, and it cleans
up the submission part.
Suggested-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There is no operation to order with afterwards, and removing the flag is
not critical in any way.
There will always be a "race condition" where the application will
trigger IORING_ENTER_SQ_WAKEUP when it isn't actually needed.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
smp_store_release in io_commit_sqring already orders the store to
dropped before the update to SQ head.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The memory operations before reading cq head are unrelated and we
don't care about their order.
Document that the control dependency in combination with READ_ONCE and
WRITE_ONCE forms a barrier we need.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
wq_has_sleeper has a full barrier internally. The smp_rmb barrier in
io_uring_poll synchronizes with it.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The application reading the CQ ring needs a barrier to pair with the
smp_store_release in io_commit_cqring, not the barrier after it.
Also a write barrier *after* writing something (but not *before*
writing anything interesting) doesn't order anything, so an smp_wmb()
after writing SQ tail is not needed.
Additionally consider reading SQ head and writing CQ tail in the notes.
Also add some clarifications how the various other fields in the ring
buffers are used.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Not all request types set REQ_F_FORCE_NONBLOCK when they needed async
punting; reverse logic instead and set REQ_F_NOWAIT if request mustn't
be punted.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Merged with my previous patch for this.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since commit 09bb839434 we don't use the state argument for any sort
of on-stack caching in the io read and write path. Remove the stale
and unused argument from them, and bubble it up to __io_submit_sqe()
and down to io_prep_rw().
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring_poll shouldn't signal EPOLLOUT | EPOLLWRNORM if the queue is
full; the old check would always signal EPOLLOUT | EPOLLWRNORM (unless
there were U32_MAX - 1 entries in the SQ queue).
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Reading the SQ tail needs to come after setting IORING_SQ_NEED_WAKEUP in
flags; there is no cheap barrier for ordering a store before a load, a
full memory barrier is required.
Userspace needs a full memory barrier between updating SQ tail and
checking for the IORING_SQ_NEED_WAKEUP too.
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A read memory barrier is required between reading SQ tail and reading
the actual data belonging to the SQ entry.
Userspace needs a matching write barrier between writing SQ entries and
updating SQ tail (using smp_store_release to update tail will do).
Signed-off-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we have multiple threads doing io_uring_register(2) on an io_uring
fd, then we can potentially try and kill the percpu reference while
someone else has already killed it.
Prevent this race by failing io_uring_register(2) if the ref is marked
dying. This is safe since we're inside the io_uring mutex.
Fixes: b19062a567 ("io_uring: fix possible deadlock between io_uring_{enter,register}")
Reported-by: syzbot <syzbot+10d25e23199614b7721f@syzkaller.appspotmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is a leftover from when the rings initially were not free flowing,
and hence a test for tail + 1 == head would indicate full. Since we now
let them wrap instead of mask them with the size, we need to check if
they drift more than the ring size from each other.
This fixes a case where we'd overwrite CQ ring entries, if the user
failed to reap completions. Both cases would ultimately result in lost
completions as the application violated the depth it asked for. The only
difference is that before this fix we'd return invalid entries for the
overflowed completions, instead of properly flagging it in the
cq_ring->overflow variable.
Reported-by: Stefan Bühler <source@stbuehler.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we have multiple threads, one doing io_uring_enter() while the other
is doing io_uring_register(), we can run into a deadlock between the
two. io_uring_register() must wait for existing users of the io_uring
instance to exit. But it does so while holding the io_uring mutex.
Callers of io_uring_enter() may need this mutex to make progress (and
eventually exit). If we wait for users to exit in io_uring_register(),
we can't do so with the io_uring mutex held without potentially risking
a deadlock.
Drop the io_uring mutex while waiting for existing callers to exit. This
is safe and guaranteed to make forward progress, since we already killed
the percpu ref before doing so. Hence later callers of io_uring_enter()
will be rejected.
Reported-by: syzbot+16dc03452dee970a0c3e@syzkaller.appspotmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since the fget/fput handling was reworked in commit 09bb839434, we
never call io_file_put() with state == NULL (and hence file != NULL)
anymore. Remove that case.
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently call cpu_possible() even if we don't use the CPU. Move the
test under the SQ_AFF branch, which is the only place where we'll use
the value. Do the cpu_possible() test AFTER we've limited it to a max
of NR_CPUS. This avoids triggering the following warning:
WARNING: CPU: 1 PID: 7600 at include/linux/cpumask.h:121 cpu_max_bits_warn
if CONFIG_DEBUG_PER_CPU_MAPS is enabled.
While in there, also move the SQ thread idle period assignment inside
SETUP_SQPOLL, as we don't use it otherwise either.
Reported-by: syzbot+cd714a07c6de2bc34293@syzkaller.appspotmail.com
Fixes: 6c271ce2f1 ("io_uring: add submission polling")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This options spawns a kernel side thread that will poll for submissions
(and completions, if IORING_SETUP_IOPOLL is set). As this allows a user
to potentially use more cycles outside of the normal hierarchy,
restrict the use of this feature to root.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Will Deacon reported the following KASAN complaint:
[ 149.890370] ==================================================================
[ 149.891266] BUG: KASAN: double-free or invalid-free in io_sqe_files_unregister+0xa8/0x140
[ 149.892218]
[ 149.892411] CPU: 113 PID: 3974 Comm: io_uring_regist Tainted: G B 5.1.0-rc3-00012-g40b114779944 #3
[ 149.893623] Hardware name: linux,dummy-virt (DT)
[ 149.894169] Call trace:
[ 149.894539] dump_backtrace+0x0/0x228
[ 149.895172] show_stack+0x14/0x20
[ 149.895747] dump_stack+0xe8/0x124
[ 149.896335] print_address_description+0x60/0x258
[ 149.897148] kasan_report_invalid_free+0x78/0xb8
[ 149.897936] __kasan_slab_free+0x1fc/0x228
[ 149.898641] kasan_slab_free+0x10/0x18
[ 149.899283] kfree+0x70/0x1f8
[ 149.899798] io_sqe_files_unregister+0xa8/0x140
[ 149.900574] io_ring_ctx_wait_and_kill+0x190/0x3c0
[ 149.901402] io_uring_release+0x2c/0x48
[ 149.902068] __fput+0x18c/0x510
[ 149.902612] ____fput+0xc/0x18
[ 149.903146] task_work_run+0xf0/0x148
[ 149.903778] do_notify_resume+0x554/0x748
[ 149.904467] work_pending+0x8/0x10
[ 149.905060]
[ 149.905331] Allocated by task 3974:
[ 149.905934] __kasan_kmalloc.isra.0.part.1+0x48/0xf8
[ 149.906786] __kasan_kmalloc.isra.0+0xb8/0xd8
[ 149.907531] kasan_kmalloc+0xc/0x18
[ 149.908134] __kmalloc+0x168/0x248
[ 149.908724] __arm64_sys_io_uring_register+0x2b8/0x15a8
[ 149.909622] el0_svc_common+0x100/0x258
[ 149.910281] el0_svc_handler+0x48/0xc0
[ 149.910928] el0_svc+0x8/0xc
[ 149.911425]
[ 149.911696] Freed by task 3974:
[ 149.912242] __kasan_slab_free+0x114/0x228
[ 149.912955] kasan_slab_free+0x10/0x18
[ 149.913602] kfree+0x70/0x1f8
[ 149.914118] __arm64_sys_io_uring_register+0xc2c/0x15a8
[ 149.915009] el0_svc_common+0x100/0x258
[ 149.915670] el0_svc_handler+0x48/0xc0
[ 149.916317] el0_svc+0x8/0xc
[ 149.916817]
[ 149.917101] The buggy address belongs to the object at ffff8004ce07ed00
[ 149.917101] which belongs to the cache kmalloc-128 of size 128
[ 149.919197] The buggy address is located 0 bytes inside of
[ 149.919197] 128-byte region [ffff8004ce07ed00, ffff8004ce07ed80)
[ 149.921142] The buggy address belongs to the page:
[ 149.921953] page:ffff7e0013381f00 count:1 mapcount:0 mapping:ffff800503417c00 index:0x0 compound_mapcount: 0
[ 149.923595] flags: 0x1ffff00000010200(slab|head)
[ 149.924388] raw: 1ffff00000010200 dead000000000100 dead000000000200 ffff800503417c00
[ 149.925706] raw: 0000000000000000 0000000080400040 00000001ffffffff 0000000000000000
[ 149.927011] page dumped because: kasan: bad access detected
[ 149.927956]
[ 149.928224] Memory state around the buggy address:
[ 149.929054] ffff8004ce07ec00: 00 00 00 00 00 00 00 00 fc fc fc fc fc fc fc fc
[ 149.930274] ffff8004ce07ec80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
[ 149.931494] >ffff8004ce07ed00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
[ 149.932712] ^
[ 149.933281] ffff8004ce07ed80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
[ 149.934508] ffff8004ce07ee00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
[ 149.935725] ==================================================================
which is due to a failure in registrering a fileset. This frees the
ctx->user_files pointer, but doesn't clear it. When the io_uring
instance is later freed through the normal channels, we free this
pointer again. At this point it's invalid.
Ensure we clear the pointer when we free it for the error case.
Reported-by: Will Deacon <will.deacon@arm.com>
Tested-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In case of direct write -EAGAIN will be returned if page cache was
previously populated. To avoid immediate completion of a request
with -EAGAIN error write has to be offloaded to the async worker,
like io_read() does.
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: linux-block@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
On big-endian architectures, the signal masks are differnet
between 32-bit and 64-bit tasks, so we have to use a different
function for reading them from user space.
io_cqring_wait() initially got this wrong, and always interprets
this as a native structure. This is ok on x86 and most arm64,
but not on s390, ppc64be, mips64be, sparc64 and parisc.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----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=l4lx
-----END PGP SIGNATURE-----
Merge tag 'io_uring-20190323' of git://git.kernel.dk/linux-block
Pull io_uring fixes and improvements from Jens Axboe:
"The first five in this series are heavily inspired by the work Al did
on the aio side to fix the races there.
The last two re-introduce a feature that was in io_uring before it got
merged, but which I pulled since we didn't have a good way to have
BVEC iters that already have a stable reference. These aren't
necessarily related to block, it's just how io_uring pins fixed
buffers"
* tag 'io_uring-20190323' of git://git.kernel.dk/linux-block:
block: add BIO_NO_PAGE_REF flag
iov_iter: add ITER_BVEC_FLAG_NO_REF flag
io_uring: mark me as the maintainer
io_uring: retry bulk slab allocs as single allocs
io_uring: fix poll races
io_uring: fix fget/fput handling
io_uring: add prepped flag
io_uring: make io_read/write return an integer
io_uring: use regular request ref counts
For ITER_BVEC, if we're holding on to kernel pages, the caller
doesn't need to grab a reference to the bvec pages, and drop that
same reference on IO completion. This is essentially safe for any
ITER_BVEC, but some use cases end up reusing pages and uncondtionally
dropping a page reference on completion. And example of that is
sendfile(2), that ends up being a splice_in + splice_out on the
pipe pages.
Add a flag that tells us it's fine to not grab a page reference
to the bvec pages, since that caller knows not to drop a reference
when it's done with the pages.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
I've seen cases where bulk alloc fails, since the bulk alloc API
is all-or-nothing - either we get the number we ask for, or it
returns 0 as number of entries.
If we fail a batch bulk alloc, retry a "normal" kmem_cache_alloc()
and just use that instead of failing with -EAGAIN.
While in there, ensure we use GFP_KERNEL. That was an oversight in
the original code, when we switched away from GFP_ATOMIC.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is a straight port of Al's fix for the aio poll implementation,
since the io_uring version is heavily based on that. The below
description is almost straight from that patch, just modified to
fit the io_uring situation.
io_poll() has to cope with several unpleasant problems:
* requests that might stay around indefinitely need to
be made visible for io_cancel(2); that must not be done to
a request already completed, though.
* in cases when ->poll() has placed us on a waitqueue,
wakeup might have happened (and request completed) before ->poll()
returns.
* worse, in some early wakeup cases request might end
up re-added into the queue later - we can't treat "woken up and
currently not in the queue" as "it's not going to stick around
indefinitely"
* ... moreover, ->poll() might have decided not to
put it on any queues to start with, and that needs to be distinguished
from the previous case
* ->poll() might have tried to put us on more than one queue.
Only the first will succeed for io poll, so we might end up missing
wakeups. OTOH, we might very well notice that only after the
wakeup hits and request gets completed (all before ->poll() gets
around to the second poll_wait()). In that case it's too late to
decide that we have an error.
req->woken was an attempt to deal with that. Unfortunately, it was
broken. What we need to keep track of is not that wakeup has happened -
the thing might come back after that. It's that async reference is
already gone and won't come back, so we can't (and needn't) put the
request on the list of cancellables.
The easiest case is "request hadn't been put on any waitqueues"; we
can tell by seeing NULL apt.head, and in that case there won't be
anything async. We should either complete the request ourselves
(if vfs_poll() reports anything of interest) or return an error.
In all other cases we get exclusion with wakeups by grabbing the
queue lock.
If request is currently on queue and we have something interesting
from vfs_poll(), we can steal it and complete the request ourselves.
If it's on queue and vfs_poll() has not reported anything interesting,
we either put it on the cancellable list, or, if we know that it
hadn't been put on all queues ->poll() wanted it on, we steal it and
return an error.
If it's _not_ on queue, it's either been already dealt with (in which
case we do nothing), or there's io_poll_complete_work() about to be
executed. In that case we either put it on the cancellable list,
or, if we know it hadn't been put on all queues ->poll() wanted it on,
simulate what cancel would've done.
Fixes: 221c5eb233 ("io_uring: add support for IORING_OP_POLL")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
