On 4/16/24 04:29, Paul Moore wrote:
On Mon, Apr 15, 2024 at 7:26 PM Dan Clash <daclash@xxxxxxxxxxxxxxxxxxx> wrote:
Below is a test program that causes multiple io_uring worker threads to
hit a NULL dereference while executing openat ops.
The test program hangs forever in a D state. The test program can be
run again after the NULL dereferences. However, there are long delays
at reboot time because the io_uring_cancel() during do_exit() attempts
to wake the worker threads.
The test program is single threaded but it queues multiple openat and
close ops with IOSQE_ASYNC set before waiting for completions.
I collected trace with /sys/kernel/tracing/events/io_uring/enable
enabled if that is helpful.
The test program reproduces similar problems in the following releases.
mainline v6.9-rc3
stable 6.8.5
Ubuntu 6.5.0-1018-azure
The test program does not reproduce the problem in Ubuntu
5.15.0-1052-azure, which does not have the io_uring audit changes.
The following is the first io_uring worker thread backtrace in the repro
against v6.9-rc3.
BUG: kernel NULL pointer dereference, address: 0000000000000000
#PF: supervisor read access in kernel mode
#PF: error_code(0x0000) - not-present page
PGD 0 P4D 0
Oops: 0000 [#1] SMP PTI
CPU: 0 PID: 4628 Comm: iou-wrk-4605 Not tainted 6.9.0-rc3 #2
Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine,
BIOS Hyper-V UEFI Release v4.1 11/28/2023
RIP: 0010:strlen (lib/string.c:402)
Call Trace:
<TASK>
? show_regs (arch/x86/kernel/dumpstack.c:479)
? __die (arch/x86/kernel/dumpstack.c:421 arch/x86/kernel/dumpstack.c:434)
? page_fault_oops (arch/x86/mm/fault.c:713)
? do_user_addr_fault (arch/x86/mm/fault.c:1261)
? exc_page_fault (./arch/x86/include/asm/irqflags.h:37
./arch/x86/include/asm/irqflags.h:72 arch/x86/mm/fault.c:1513
arch/x86/mm/fault.c:1563)
? asm_exc_page_fault (./arch/x86/include/asm/idtentry.h:623)
? __pfx_strlen (lib/string.c:402)
? parent_len (kernel/auditfilter.c:1284).
__audit_inode (kernel/auditsc.c:2381 (discriminator 4))
Thanks for the well documented bug report!
That's interesting, it looks like audit_inode() is potentially being
passed a filename struct with a NULL name field (filename::name ==
NULL). Given the IOSQE_ASYNC and what looks like io_uring calling
getname() from within the __io_openat_prep() function, I suspect the
issue is that we aren't associating the filename information we
collect in getname() with the proper audit_context(). In other words,
we do the getname() in one context, and then the actual open operation
in another, and the audit filename info is lost in the switch.
I think this is related to another issue that Jens and I have been
discussing relating to connect() and sockaddrs. We had been hoping
that the issue we were seeing with sockaddrs was just a special case
with connect, but it doesn't look like that is the case.
I'm going to be a bit busy this week with conferences, but given the
previous discussions with Jens as well as this new issue, I suspect
that we are going to need to do some work to support creation of a
thin, or lazily setup, audit_context that we can initialize in the
io_uring prep routines for use by getname(), move_addr_to_kernel(),
etc., store in the io_kiocb struct, and then fully setup in
audit_uring_entry().
I'd prefer not to leak that much into the io_uring's hot path. I
don't understand specifics of the problem, but let me throw
some ideas:
1) Each io_uring request has a reference to the task it was
submitted from, i.e. req->task, can you use the context from
the submitter task? E.g.
audit_ctx = req->task->audit_context
io_uring explicitly lists all tasks using it, and you can easily
hook in there and initialise audit so that req->ctx->audit_context
is always set.
2) Can we initialise audit for each io-wq task when we create
them? We can also try to share audit ctx b/w iowq tasks and
the task they were created for.
? link_path_walk.part.0.constprop.0 (fs/namei.c:2324)
path_openat (fs/namei.c:3550 fs/namei.c:3796)
do_filp_open (fs/namei.c:3826)
? alloc_fd (./arch/x86/include/asm/paravirt.h:589 (discriminator 10)
./arch/x86/include/asm/qspinlock.h:57 (discriminator 10)
./include/linux/spinlock.h:204 (discriminator 10)
./include/linux/spinlock_api_smp.h:142 (discriminator 10)
./include/linux/spinlock.h:391 (discriminator 10) fs/file.c:553
(discriminator 10))
io_openat2 (io_uring/openclose.c:140)
io_openat (io_uring/openclose.c:178)
io_issue_sqe (io_uring/io_uring.c:1897)
io_wq_submit_work (io_uring/io_uring.c:2006)
io_worker_handle_work (io_uring/io-wq.c:540 io_uring/io-wq.c:597)
io_wq_worker (io_uring/io-wq.c:258 io_uring/io-wq.c:648)
? __pfx_io_wq_worker (io_uring/io-wq.c:627)
? raw_spin_rq_unlock (./arch/x86/include/asm/paravirt.h:589
./arch/x86/include/asm/qspinlock.h:57 ./include/linux/spinlock.h:204
./include/linux/spinlock_api_smp.h:142 kernel/sched/core.c:603)
? finish_task_switch.isra.0 (./arch/x86/include/asm/irqflags.h:42
./arch/x86/include/asm/irqflags.h:77 kernel/sched/sched.h:1397
kernel/sched/core.c:5163 kernel/sched/core.c:5281)
? __pfx_io_wq_worker (io_uring/io-wq.c:627)
ret_from_fork (arch/x86/kernel/process.c:156)
? __pfx_io_wq_worker (io_uring/io-wq.c:627)
ret_from_fork_asm (arch/x86/entry/entry_64.S:256)
--
Pavel Begunkov
