Re: [LSF/MM/BPF BoF]: extend UBLK to cover real storage hardware

On Thu, Feb 23, 2023 at 07:17:33AM +0800, Ming Lei wrote:

On Sat, Feb 18, 2023 at 01:38:08PM -0500, Stefan Hajnoczi wrote:

quoted

On Sat, Feb 18, 2023 at 07:22:49PM +0800, Ming Lei wrote:

quoted

On Fri, Feb 17, 2023 at 11:39:58AM -0500, Stefan Hajnoczi wrote:

quoted

On Fri, Feb 17, 2023 at 10:20:45AM +0800, Ming Lei wrote:

quoted

On Thu, Feb 16, 2023 at 12:21:32PM +0100, Andreas Hindborg wrote:

quoted

Ming Lei [off-list ref] writes:

quoted

On Thu, Feb 16, 2023 at 10:44:02AM +0100, Andreas Hindborg wrote:

quoted

Hi Ming,

Ming Lei [off-list ref] writes:

quoted

On Mon, Feb 13, 2023 at 02:13:59PM -0500, Stefan Hajnoczi wrote:

quoted

On Mon, Feb 13, 2023 at 11:47:31AM +0800, Ming Lei wrote:

quoted

On Wed, Feb 08, 2023 at 07:17:10AM -0500, Stefan Hajnoczi wrote:

quoted

On Wed, Feb 08, 2023 at 10:12:19AM +0800, Ming Lei wrote:

quoted

On Mon, Feb 06, 2023 at 03:27:09PM -0500, Stefan Hajnoczi wrote:

quoted

On Mon, Feb 06, 2023 at 11:00:27PM +0800, Ming Lei wrote:

quoted

Hello,

So far UBLK is only used for implementing virtual block device from
userspace, such as loop, nbd, qcow2, ...[1].

I won't be at LSF/MM so here are my thoughts:

Thanks for the thoughts, :-)

quoted

quoted

It could be useful for UBLK to cover real storage hardware too:

- for fast prototype or performance evaluation

- some network storages are attached to host, such as iscsi and nvme-tcp,
the current UBLK interface doesn't support such devices, since it needs
all LUNs/Namespaces to share host resources(such as tag)

Can you explain this in more detail? It seems like an iSCSI or
NVMe-over-TCP initiator could be implemented as a ublk server today.
What am I missing?

The current ublk can't do that yet, because the interface doesn't
support multiple ublk disks sharing single host, which is exactly
the case of scsi and nvme.

Can you give an example that shows exactly where a problem is hit?

I took a quick look at the ublk source code and didn't spot a place
where it prevents a single ublk server process from handling multiple
devices.

Regarding "host resources(such as tag)", can the ublk server deal with
that in userspace? The Linux block layer doesn't have the concept of a
"host", that would come in at the SCSI/NVMe level that's implemented in
userspace.

I don't understand yet...

blk_mq_tag_set is embedded into driver host structure, and referred by queue
via q->tag_set, both scsi and nvme allocates tag in host/queue wide,
that said all LUNs/NSs share host/queue tags, current every ublk
device is independent, and can't shard tags.

Does this actually prevent ublk servers with multiple ublk devices or is
it just sub-optimal?

It is former, ublk can't support multiple devices which share single host
because duplicated tag can be seen in host side, then io is failed.

I have trouble following this discussion. Why can we not handle multiple
block devices in a single ublk user space process?

From this conversation it seems that the limiting factor is allocation
of the tag set of the virtual device in the kernel? But as far as I can
tell, the tag sets are allocated per virtual block device in
`ublk_ctrl_add_dev()`?

It seems to me that a single ublk user space process shuld be able to
connect to multiple storage devices (for instance nvme-of) and then
create a ublk device for each namespace, all from a single ublk process.

Could you elaborate on why this is not possible?

If the multiple storages devices are independent, the current ublk can
handle them just fine.

But if these storage devices(such as luns in iscsi, or NSs in nvme-tcp)
share single host, and use host-wide tagset, the current interface can't
work as expected, because tags is shared among all these devices. The
current ublk interface needs to be extended for covering this case.

Thanks for clarifying, that is very helpful.

Follow up question: What would the implications be if one tried to
expose (through ublk) each nvme namespace of an nvme-of controller with
an independent tag set?

https://lore.kernel.org/linux-block/877cwhrgul.fsf@metaspace.dk/T/#m57158db9f0108e529d8d62d1d56652c52e9e3e67 (local)

quoted

What are the benefits of sharing a tagset across
all namespaces of a controller?

The userspace implementation can be simplified a lot since generic
shared tag allocation isn't needed, meantime with good performance
(shared tags allocation in SMP is one hard problem)

In NVMe, tags are per Submission Queue. AFAIK there's no such thing as
shared tags across multiple SQs in NVMe. So userspace doesn't need an

In reality the max supported nr_queues of nvme is often much less than
nr_cpu_ids, for example, lots of nvme-pci devices just support at most
32 queues, I remembered that Azure nvme supports less(just 8 queues).
That is because queue isn't free in both software and hardware, which
implementation is often tradeoff between performance and cost.

I didn't say that the ublk server should have nr_cpu_ids threads. I
thought the idea was the ublk server creates as many threads as it needs
(e.g. max 8 if the Azure NVMe device only has 8 queues).

Do you expect ublk servers to have nr_cpu_ids threads in all/most cases?

No.

In ublksrv project, each pthread maps to one unique hardware queue, so total
number of pthread is equal to nr_hw_queues.

Good, I think we agree on that part.

Here is a summary of the ublk server model I've been describing:
1. Each pthread has a separate io_uring context.
2. Each pthread has its own hardware submission queue (NVMe SQ, SCSI
   command queue, etc).
3. Each pthread has a distinct subrange of the tag space if the tag
   space is shared across hardware submission queues.
4. Each pthread allocates tags from its subrange without coordinating
   with other threads. This is cheap and simple.
5. When the pthread runs out of tags it either suspends processing new
   ublk requests or enqueues them internally. When hardware completes
   requests, the pthread resumes requests that were waiting for tags.

This way multiple ublk_devices can be handled by a single ublk server
without the Linux block layer knowing the exact tag space sharing
relationship between ublk_devices and hardware submission queues (NVMe
SQ, SCSI command queue, etc).

When ublk adds support for configuring tagsets, then 3, 4, and 5 can be
eliminated. However, this is purely an optimization. Not that much
userspace code will be eliminated and the performance gain is not huge.

I believe this model works for the major storage protocols like NVMe and
SCSI.

I put forward this model to explain why I don't agree that ublk doesn't
support ublk servers with multiple devices (e.g. I/O would be failed due
to duplicated tags).

I think we agree on 1 and 2. It's 3, 4, and 5 that I think you are
either saying won't work or are very complex/hard?

quoted

quoted

Not mention, most of scsi devices are SQ in which tag allocations from
all CPUs are against single shared tagset.

So there is still per-queue tag allocations from different CPUs which aims
at same queue.

What we discussed are supposed to be generic solution, not something just
for ideal 1:1 mapping device, which isn't dominant in reality.

The same trivial tag allocation can be used for SCSI: instead of a
private tag namespace (e.g. 0x0-0xffff), give each queue a private
subset of the tag namespace (e.g. queue 0 has 0x0-0x7f, queue 1 has
0x80-0xff, etc).

Sorry, I may not get your point.

Each hw queue has its own tag space, for example, one scsi adaptor has 2
queues, queue depth is 128, then each hardware queue's tag space is
0 ~ 127.

Also if there are two LUNs attached to this host, the two luns
share the two queue's tag space, that means any IO issued to queue 0,
no matter if it is from lun0 or lun1, the allocated tag has to unique in
the set of 0~127.

I'm trying to explain why tag allocation in userspace is simple and
cheap thanks to the ublk server's ability to create only as many threads
as hardware queues (e.g. NVMe SQs).

Even in the case where all hardware (NVME/SCSI/etc) queues and LUNs
share the same tag space (the worst case), ublk server threads can
perform allocation from distinct subranges of the shared tag space.
There are no SMP concerns because there is no overlap in the tag space
between threads.

quoted

The issue is not whether the tag namespace is shared across queues, but
the threading model of the ublk server. If the threading model requires
queues to be shared, then it becomes more complex and slow.

ublksrv's threading model is simple, each thread handles IOs from one unique
hw queue, so total thread number is equal to nr_hw_queues.

Here "hw queue" is a Linux block layer hw queue, not a hardware queue
(i.e. NVMe SQ)?

If nr_hw_queues(nr_pthreads) < nr_cpu_id, one queue(ublk pthread) has to
handle IO requests from more than one CPUs, then contention on tag allocation
from this queue(ublk pthread).

Userspace doesn't need to worry about the fact that I/O requests were
submitted by many CPUs. Each pthread processes one ublk_queue with a
known queue depth.

Each pthread has a range of userspace tags available and if there are no
more tags available then it waits to complete in-flight I/O before
accepting more requests or it can internally queue incoming requests.

quoted

It's not clear to me why you think ublk servers should choose threading
models that require queues to be shared? They don't have to. Unlike the
kernel, they can choose the number of threads.

queue sharing or not simply depends on if nr_hw_queues is less than
nr_cpu_id. That is one easy math problem, isn't it?

We're talking about different things. I mean sharing a hardware queue
(i.e. NVMe SQ) across multiple ublk server threads. You seem to define
queue sharing as multiple CPUs submitting I/O via ublk?

Thinking about your scenario: why does it matter if multiple CPUs submit
I/O to a single ublk_queue? I don't see how it makes a difference
whether 1 CPU or multiple CPUs enqueue requests on a single ublk_queue.
Userspace will process that ublk_queue in the same way in either case.

quoted

quoted

quoted

SMP tag allocator in the first place:
- Each ublk server thread has a separate io_uring context.
- Each ublk server thread has its own NVMe Submission Queue.
- Therefore it's trivial and cheap to allocate NVMe CIDs in userspace
  because there are no SMP concerns.

It isn't even trivial for 1:1 mapping, when any ublk server crashes
global tag will be leaked, and other ublk servers can't use the
leaked tag any more.

I'm not sure what you're describing here, a multi-process ublk server?
Are you saying userspace must not do tag allocation itself because it
won't be able to recover?

No matter if the ublk server is multi process or threads. If tag
allocation is implemented in userspace, you have to take thread/process
panic into account. Because if one process/pthread panics without
releasing one tag, the tag won't be visible to other ublk server any
more.

That is because each queue's tag space is shared for all LUNs/NSs which
are supposed to implemented as ublk server.

Tag utilization highly affects performance, and recover could take a
bit long or even not recovered, during this period, the leaked tags
aren't visible for other LUNs/NSs(ublk server), not mention for fixing
tag leak in recover, you have to track each tag's user(ublk server info),
which adds cost/complexity to fast/parallel io path, trivial to solve?

In the interest of time, let's defer the recovery discussion until
after the core discussion is finished. I would need to research how ublk
recovery works. I am happy to do that if you think recovery is the
reason why userspace cannot allocate tags, but reaching a conclusion on
the core discussion might be enough to make discussing recovery
unnecessary.

Thanks,
Stefan