On 8/3/22 1:28 PM, Paul Moore wrote:
> On Wed, Aug 3, 2022 at 9:16 AM Paul Moore <paul@xxxxxxxxxxxxxx> wrote:
>> On Wed, Aug 3, 2022 at 1:03 AM Peilin Ye <yepeilin.cs@xxxxxxxxx> wrote:
>>>
>>> Currently @audit_context is allocated twice for io_uring workers:
>>>
>>> 1. copy_process() calls audit_alloc();
>>> 2. io_sq_thread() or io_wqe_worker() calls audit_alloc_kernel() (which
>>> is effectively audit_alloc()) and overwrites @audit_context,
>>> causing:
>>>
>>> BUG: memory leak
>>> unreferenced object 0xffff888144547400 (size 1024):
>>> <...>
>>> hex dump (first 32 bytes):
>>> 00 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 ................
>>> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
>>> backtrace:
>>> [<ffffffff8135cfc3>] audit_alloc+0x133/0x210
>>> [<ffffffff81239e63>] copy_process+0xcd3/0x2340
>>> [<ffffffff8123b5f3>] create_io_thread+0x63/0x90
>>> [<ffffffff81686604>] create_io_worker+0xb4/0x230
>>> [<ffffffff81686f68>] io_wqe_enqueue+0x248/0x3b0
>>> [<ffffffff8167663a>] io_queue_iowq+0xba/0x200
>>> [<ffffffff816768b3>] io_queue_async+0x113/0x180
>>> [<ffffffff816840df>] io_req_task_submit+0x18f/0x1a0
>>> [<ffffffff816841cd>] io_apoll_task_func+0xdd/0x120
>>> [<ffffffff8167d49f>] tctx_task_work+0x11f/0x570
>>> [<ffffffff81272c4e>] task_work_run+0x7e/0xc0
>>> [<ffffffff8125a688>] get_signal+0xc18/0xf10
>>> [<ffffffff8111645b>] arch_do_signal_or_restart+0x2b/0x730
>>> [<ffffffff812ea44e>] exit_to_user_mode_prepare+0x5e/0x180
>>> [<ffffffff844ae1b2>] syscall_exit_to_user_mode+0x12/0x20
>>> [<ffffffff844a7e80>] do_syscall_64+0x40/0x80
>>>
>>> Then,
>>>
>>> 3. io_sq_thread() or io_wqe_worker() frees @audit_context using
>>> audit_free();
>>> 4. do_exit() eventually calls audit_free() again, which is okay
>>> because audit_free() does a NULL check.
>>>
>>> Free the old @audit_context first in audit_alloc_kernel(), and delete
>>> the redundant calls to audit_free() for less confusion.
>>>
>>> Fixes: 5bd2182d58e9 ("audit,io_uring,io-wq: add some basic audit support to io_uring")
>>> Cc: stable@xxxxxxxxxxxxxxx
>>> Signed-off-by: Peilin Ye <peilin.ye@xxxxxxxxxxxxx>
>>> ---
>>> Hi all,
>>>
>>> A better way to fix this memleak would probably be checking
>>> @args->io_thread in copy_process()? Something like:
>>>
>>> if (args->io_thread)
>>> retval = audit_alloc_kernel();
>>> else
>>> retval = audit_alloc();
>>>
>>> But I didn't want to add another if to copy_process() for this bugfix.
>>> Please suggest, thanks!
>>
>> Thanks for the report and patch! I'll take a closer look at this
>> today and get back to you.
>
> I think the best solution to this is simply to remove the calls to
> audit_alloc_kernel() in the io_uring and io-wq code, as well as the
> audit_alloc_kernel() function itself. As long as create_io_thread()
> ends up calling copy_process to create the new kernel thread the
> audit_context should be allocated correctly. Peilin Ye, are you able
> to draft a patch to do that and give it a test?
>
> For those that may be wondering how this happened (I definitely was!),
> it looks like when I first started working on the LSM/audit support
> for io_uring it was before the v5.12-rc1 release when
> create_io_thread() was introduced. Prior to create_io_thread() it
> appears that io_uring/io-wq wasn't calling into copy_process() and
> thus was not getting an audit_context allocated in the kernel thread's
> task_struct; the solution for those original development drafts was to
> add a call to a new audit_alloc_kernel() which would handle the
> audit_context allocation. Unfortunately, I didn't notice the move to
> create_io_thread() during development and the redundant
> audit_alloc_kernel() calls remained :/
I agree with your analysis and suggested solution. Post the native io-wq
workers create_io_thread() -> copy_process() is always used for io-wq
(and sqpoll, for that matter).
--
Jens Axboe
