[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-block] BDRV request fragmentation and virtio-blk write submiss

From: Евгений Яковлев
Subject: Re: [Qemu-block] BDRV request fragmentation and virtio-blk write submission guarantees (2nd try)
Date: Thu, 18 Jul 2019 18:27:40 +0300
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.7.2

Evgeny Yakovlev
Lead Software Engineer, Yandex.Cloud Hypervisor Team
On 18.07.2019 17:59, Kevin Wolf wrote:
Am 18.07.2019 um 15:52 hat Евгений Яковлев geschrieben:
Hi everyone,

My previous message was misformatted, so here's another one. Sorry about

We're currently working on implementing a qemu BDRV format driver which we
are using with virtio-blk devices.

I have a question concerning BDRV request fragmentation and virtio-blk write
request submission which is not entirely clear to me by only reading virtio
spec. Could you please consider the following case and give some additional

1. Our BDRV format driver has a notion of max supported transfer size. So we
implement BlockDriver::bdrv_refresh_limits where we fill out
BlockLimits::max_transfer and opt_transfer fields.

2. virtio-blk exposes max_transfer as a virtio_blk_config::opt_io_size
field, which (according to spec 1.1) is a **suggested** maximum. We read
"suggested" as "guest driver may still send requests that don't fit into
opt_io_size and we should handle those"...

3. ... and judging by code in block/io.c qemu block layer handles such
requests by fragmenting them into several BDRV requests if request size is >

4. Guest will see request completion only after all fragments are handled.
However each fragment submission path can call qemu_coroutine_yield and move
on to submitting next request available in virtq before completely
submitting the rest of the fragments. Which means the following situation is
possible where BDRV sees 2 write requests in virtq, both of which are larger
than max_transfer:

Blocks: -----------------------------

Write1: ------xxxxxxxx

Write2: ------yyyyyyyy

Write1Chunk1: xxxx

Write2Chunk1: yyyy

Write2Chunk2: ----yyyy

Write1Chunk1: ----xxxx

Blocks: ------yyyyxxxx-----------------

In above scenario guest virtio-blk driver decided to submit 2 intersecting
write requests, both of which are larger than ||max_transfer, and then call

I understand that virtio-blk may handle requests out of order, so guest must
not make any assumptions on relative order in which those requests will be

However, can guest driver expect that whatever the submission order will be,
the actual intersecting writes will be atomic?

In other words, will it be correct for conforming virtio-blk driver to
expect only "xxxxxxxx" or "yyyyyyyy" but not anything else in between, after
both requests are reported as completed?

Because i think that is something that may happen in qemu right now, if i
understood correctly.
I don't think atomicity is promised anywhere in the virtio
specification, and I agree with you that this case can happen (it
probably happens much more frequently when you use image formats instead
of raw files).

On the other hand, there is no good reason for a guest OS to submit two
write request to the same blocks in parallel. Even if it could expect
that one of the requests wins, the end result would still be undefined,
so I don't think this could ever be a useful thing to do. (Well, I guess
it could replace flipping a coin...)

Thanks Kevin. I agree that described guest behavior does not a have a sensible reason behind it. However, just based on purely theoretical basis, according to virtio-blk contract, is it valid for guest to even _assume_ that above situation with 2 requests _must_ be resolved in one of two specific cases i described and not anything in between? In other words that writes will be atomic even if their relative order is undefined. We could not get a clear answer from virtio spec ourselves.

For instance, IIRC, nvme spec declares atomicity guarantees as well as ordering for specific commands ("6.4 Atomic Operations").


reply via email to

[Prev in Thread] Current Thread [Next in Thread]