Re: Memory coherency issue with IO thread offloading?

Jens Axboe [off-list ref] writes:

Hi,

Hi Jens,

Thanks for the report.

I got a report sent to me from mariadb, in where 5.10.158 works fine and
5.10.162 is broken. And in fact, current 6.3-rc also fails the test
case. Beware that this email is long, as I'm trying to include
everything that may be relevant...

The test case in question is pretty simple. On debian testing, do:

$ sudo apt-get install mariadb-test
$ cd /usr/share/mysql/mysql-test
$ ./mtr --mysqld=--innodb-flush-method=fsync --mysqld=--innodb-use-native-aio=1 --vardir=/dev/shm/mysql  --force encryption.innodb_encryption,innodb,undo0 --repeat=200

I mostly use Fedora, the package name is the same but the mtr binary
ends up in /usr/share/mysql.

and if it fails, you'll see something like:

encryption.innodb_encryption 'innodb,undo0' [ 6 pass ]   3120
encryption.innodb_encryption 'innodb,undo0' [ 7 pass ]   3123
encryption.innodb_encryption 'innodb,undo0' [ 8 pass ]   3042
encryption.innodb_encryption 'innodb,undo0' [ 9 fail ]
        Test ended at 2023-03-23 16:55:17

I haven't been able to get this to fail yet. I've done several runs with
--repeat=500 and haven't seen any errors yet.

Are there any CONFIG options I'd need to trip this?

...

Today I finally gave up and ran a basic experiment, which simply
offloads the writes to a kthread. Since powerpc has an interesting
memory coherency model, my suspicion was that the work involved with
switching MMs for the kthread could just be the main difference here.
The patch is really dumb and simple - rather than queue the write to an
IO thread, it just offloads it to a kthread that then does
kthread_use_mm(), perform write with the same write handler,
kthread_unuse_mm(). AND THIS WORKS! Usually the above mtr test would
fail in 2..20 loops, I've now done 200 and 500 loops and it's fine.

Can you share the patch that does that? It would help me track down
where exactly in the io_uring code you're talking about.

Which then leads me to the question, what about the IO thread offload
makes this fail on powerpc (and no other arch I've tested on, including
x86/x86-64/aarch64/hppa64)? The offload should be equivalent to having a
thread in userspace in the application, and having that thread just
perform the writes. Is there some magic involved with the kthread mm
use/unuse that makes this sufficiently consistent on powerpc? I've tried
any mix of isync()/mb and making the flush_dcache_page() unconditionally
done in the filemap read/write helpers, and it still falls flat on its
face with the offload to an IO thread.

My first guess would be that there's some missing barriers between the
thread that queues the IO and the IO worker thread. 

I think you're using schedule_work() for that though, which should be a
full barrier. Could it be on the completion side?

I can't think of any magic in kthread_use_mm() other than extra
barriers. In particular kthread_unuse_mm() has an
smp_mb__after_spinlock() which is a full memory barrier on powerpc but
is a nop on some other architectures, x86 at least.

I must clearly be missing something here, which is why I'm emailing the
powerpc Gods for help :-)

Unfortunately the true God of powerpc memory ordering has left us and
ascended into the Metaverse ;)

cheers