Re: [PATCH] filesetup: do not ask O_RDWR for read-only workload

[Jens Axboe <axboe@kernel.dk>]

On 2/7/23 9:16?AM, Vincent Fu wrote:
> On 2/6/23 14:37, Jens Axboe wrote:
>> On 2/5/23 11:52?PM, Kanchan Joshi wrote:
>>> On Fri, Feb 03, 2023 at 10:26:15AM -0500, Vincent Fu wrote:
>>>> On 2/3/23 09:14, Jens Axboe wrote:
>>>>> On 2/3/23 7:08?AM, Vincent Fu wrote:
>>>>>> On 2/3/23 07:46, Jens Axboe wrote:
>>>>>>> On Feb 3, 2023, at 5:35 AM, Kanchan Joshi <joshi.k@samsung.com> wrote:
>>>>>>>>
>>>>>>>> ?Use O_RDONLY flag when read is requested on char-type files.
>>>>>>>>
>>>>>>>> Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
>>>>>>>> ---
>>>>>>>>
>>>>>>>> Before this patch: unexpected permission-denial for unprivileged-user.
>>>>>>>>
>>>>>>>> $ ls -l /dev/ng0n1
>>>>>>>> cr--r--r-- 1 root root 242, 0 Feb  3 16:30 /dev/ng0n1
>>>>>>>>
>>>>>>>> $ ./fio -iodepth=1 -rw=randread -ioengine=io_uring_cmd -bs=4k -numjobs=1
>>>>>>>> -size=4k -cmd_type=nvme -filename=/dev/ng0n1 -name=t
>>>>>>>> t: (g=0): rw=randread, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T)
>>>>>>>> 4096B-4096B, ioengine=io_uring_cmd, iodepth=1
>>>>>>>> fio-3.33-71-g7d7a
>>>>>>>> Starting 1 process
>>>>>>>> fio: pid=131312, err=13/file:filesetup.c:805, func=open(/dev/ng0n1),
>>>>>>>> error=Permission denied
>>>>>>>>
>>>>>>>> filesetup.c | 5 +----
>>>>>>>> 1 file changed, 1 insertion(+), 4 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/filesetup.c b/filesetup.c
>>>>>>>> index 1d3cc5ad..d77b8ba4 100644
>>>>>>>> --- a/filesetup.c
>>>>>>>> +++ b/filesetup.c
>>>>>>>> @@ -768,10 +768,7 @@ open_again:
>>>>>>>>          else
>>>>>>>>              from_hash = file_lookup_open(f, flags);
>>>>>>>>      } else if (td_read(td)) {
>>>>>>>> -        if (f->filetype == FIO_TYPE_CHAR && !read_only)
>>>>>>>> -            flags |= O_RDWR;
>>>>>>>> -        else
>>>>>>>> -            flags |= O_RDONLY;
>>>>>>>> +        flags |= O_RDONLY;
>>>>>>>
>>>>>>> This will break sg like interfaces, where a read is done by writing the command to the char device.
>>>>>>>
>>>>>>> ?
>>>>>>> Jens Axboe
>>>>>>>
>>>>>>
>>>>>> Kanchan, does it work if you run fio with the --readonly option?
>>>>>
>>>>> I think we should just make it work. Conceptually, the patch is obviously
>>>>> fine, it just happens to break some oddball cases like sg/bsg. But maybe
>>>>> we just add a fio ioengine flag for that, like FIO_RO_NEEDS_RW_OPEN or
>>>>> something, where the engine can tell us if it needs a writeable open
>>>>> even or a read-only workload.
>>>>>
>>>>> Then generic_file_open() can use O_RDONLY for td_read(), except if the
>>>>> engine has FIO_RO_NEEDS_RW_OPEN set.
>>>>>
>>>>
>>>> Kanchan, can you try the diff below? It's also available at https://protect2.fireeye.com/v1/url?k=37173803-566c928b-3716b34c-74fe4860018a-aac6167af74310c3&q=1&e=3e678cf7-7d3f-487a-8a94-a9f0b41ca778&u=https%3A%2F%2Fgithub.com%2Fvincentkfu%2Ffio%2Ftree%2Frwopen
>>>
>>> Tried, this works, thanks!
>>> The patch looks good. Seems this flag can be added to libzbc engine too.
>>> That also treats the the char-type open in the same way.
>>
>> Thanks for testing - I've added it for libzbc as well.
>>
> 
> The new flag only affects generic_file_open() but the libzbc ioengine
> has its own file open function. So the flag has no effect.

That's a good point, not sure how I missed that when I explicitly
checked sg for it last week...

> Moreover libzbc uses the SG_IO ioctl even for character devices. So
> the code in libzbc_open_dev() (copied from generic_file_open) that
> opens character devices with O_RDWR is actually unnecessary.
> 
> I've committed some patches to tidy things up.

Thanks!

-- 
Jens Axboe