Re: Data Corruption bug with Samba's vfs_iouring and Linux 5.6.7/5.7rc3

[Jens Axboe <axboe@kernel.dk>]

On 5/5/20 10:53 AM, Jens Axboe wrote:
> On 5/5/20 9:44 AM, Jens Axboe wrote:
>> On 5/5/20 8:41 AM, Jens Axboe wrote:
>>> On 5/5/20 4:04 AM, Stefan Metzmacher wrote:
>>>> Hi Jens,
>>>>
>>>> we currently have a bug report [1][2] regarding a data corruption with
>>>> Samba's vfs_io_uring.c [3].
>>>>
>>>> Is there're a know problem in newer kernels? It seems the 5.3 Kernel
>>>> doesn't have the problem (at least Jeremy wasn't able to reproduce it
>>>> on the Ubuntu 5.3 kernel).
>>>>
>>>> Do you have any hints how to track that down?
>>>
>>> I'll take a look at this! Any chance Jeremy can try 5.4 and 5.5 as well,
>>> just to see where we're at, roughly? That might be very helpful.
>>
>> Trying to setup samba in a vm here to attempt to reproduce. I'm a major
>> samba noob, running with the smb.conf from the reporters email, I get:
>>
>> [2020/05/05 15:43:07.126674,  0] ../../source4/smbd/server.c:629(binary_smbd_main)
>>   samba version 4.12.2 started.
>>   Copyright Andrew Tridgell and the Samba Team 1992-2020
>> [2020/05/05 15:43:07.152828,  0] ../../source4/smbd/server.c:826(binary_smbd_main)
>>   At this time the 'samba' binary should only be used for either:
>>   'server role = active directory domain controller' or to access the ntvfs file server with 'server services = +smb' or the rpc proxy with 'dcerpc endpoint servers = remote'
>>   You should start smbd/nmbd/winbindd instead for domain member and standalone file server tasks
>> [2020/05/05 15:43:07.152937,  0] ../../lib/util/become_daemon.c:121(exit_daemon)
>>   exit_daemon: daemon failed to start: Samba detected misconfigured 'server role' and exited. Check logs for details, error code 22
>>
>> Clue bat appreciated.
> 
> Got it working, but apparently the arch samba doesn't come with io_uring...
> One question, though, from looking at the source:
> 
> static ssize_t vfs_io_uring_pread_recv(struct tevent_req *req,
> 				  struct vfs_aio_state *vfs_aio_state)
> {
> [...]
> 	if (state->ur.cqe.res < 0) {
> 		vfs_aio_state->error = -state->ur.cqe.res;
> 		ret = -1;
> 	} else {
> 		vfs_aio_state->error = 0;
> 		ret = state->ur.cqe.res;
> 	}
> 
> 	tevent_req_received(req);
> [...]
> 
> I'm assuming this is dealing with short reads?
> 
> I'll try and see if I can get an arch binary build that has the
> vfs_io_uring module and reproduce.

Got that done, and I can now mount it on Linux. Been trying pretty
hard to trigger any corruptions on reads, but it works for me. Checked
that we see short reads, and we do, and that it handles it just fine.
So pretty blank right now on what this could be.

FWIW, I'm mounting on Linux as:

# mount -t cifs -o ro,guest //arch/data /smb

-- 
Jens Axboe