On 11/7/20 3:49 PM, Pavel Begunkov wrote:
> On 07/11/2020 22:30, Jens Axboe wrote:
>> On 11/7/20 2:16 PM, Pavel Begunkov wrote:
>>> SQPOLL task may find sqo_task->files == NULL, so
>>> __io_sq_thread_acquire_files() would left it unset and so all the
>>> following fails, e.g. attempts to submit. Fail if sqo_task doesn't have
>>> files.
>>>
>>> [ 118.962785] BUG: kernel NULL pointer dereference, address:
>>> 0000000000000020
>>> [ 118.963812] #PF: supervisor read access in kernel mode
>>> [ 118.964534] #PF: error_code(0x0000) - not-present page
>>> [ 118.969029] RIP: 0010:__fget_files+0xb/0x80
>>> [ 119.005409] Call Trace:
>>> [ 119.005651] fget_many+0x2b/0x30
>>> [ 119.005964] io_file_get+0xcf/0x180
>>> [ 119.006315] io_submit_sqes+0x3a4/0x950
>>> [ 119.006678] ? io_double_put_req+0x43/0x70
>>> [ 119.007054] ? io_async_task_func+0xc2/0x180
>>> [ 119.007481] io_sq_thread+0x1de/0x6a0
>>> [ 119.007828] kthread+0x114/0x150
>>> [ 119.008135] ? __ia32_sys_io_uring_enter+0x3c0/0x3c0
>>> [ 119.008623] ? kthread_park+0x90/0x90
>>> [ 119.008963] ret_from_fork+0x22/0x30
>>>
>>> Reported-by: Josef Grieb <josef.grieb@xxxxxxxxx>
>>> Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
>>> ---
>>> fs/io_uring.c | 19 ++++++++++++-------
>>> 1 file changed, 12 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>> index 8d721a652d61..9c035c5c4080 100644
>>> --- a/fs/io_uring.c
>>> +++ b/fs/io_uring.c
>>> @@ -1080,7 +1080,7 @@ static void io_sq_thread_drop_mm_files(void)
>>> }
>>> }
>>>
>>> -static void __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
>>> +static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
>>> {
>>> if (!current->files) {
>>> struct files_struct *files;
>>> @@ -1091,7 +1091,7 @@ static void __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
>>> files = ctx->sqo_task->files;
>>> if (!files) {
>>> task_unlock(ctx->sqo_task);
>>> - return;
>>> + return -EFAULT;
>>
>> I don't think we should use -EFAULT here, it's generally used for trying
>> to copy in/out of invalid regions. Probably -ECANCELED is better here,
>
> Noted, I'll resend after Josef tests this.
>
>> in lieu of something super appropriate. Maybe -EBADF would be fine too.
>
> Yeah, something along OWNER_TASK_DEAD would make more sense.
You could try and commandeer -EOWNERDEAD for this use case, it does
make sense.
--
Jens Axboe
