[dm-devel] Unable to receive overwrite BIO in dm-thin

[Mike Snitzer]

On Mon, Sep 23 2013 at  7:06am -0400,
Teng-Feng Yang <shinrairis at gmail.com> wrote:

> Hi folks,
> 
> I have recently performed some experiments to get the IO performance
> of thin devices created by dm-thin under different circumstances.
> Therefore, I create a 100GB thin device from a thin pool (block size =
> 1MB) created by a 3TB HD as the data device and a 128GB SSD as the
> metadata device.
> 
> First, I want to know the IO performance of the raw HD
> 
> > dd if=/dev/zero of=/dev/sdg bs=1M count=10K
> 10240+0 records in
> 10240+0 records out
> 10737418240 bytes (11 GB) copied, 79.3541 s, 135 MB/s
> 
> Then, I create a thin device and do the same IO.
> 
> > dd if=/dev/zero of=/dev/mapper/thin bs=1M count=10K
> 1024+0 records in
> 1024+0 records out
> 1073741824 bytes (1.1 GB) copied, 22.4915 s, 47.7 MB/s
> 
> The write throughput is much more lower than the raw device, so I dig
> a little deeper into the source code and turn the block_dump flag to
> true.
> It turns out that the "max_sectors_kb" of the thin device has been set
> to 1024 sectors ( 512KB ). so the thin device can never receive 1MB
> block size IO and try to zero block before every write.
> So, I remove the whole pool and recreate the whole testing environment
> and then set the max_sectors_kb to 2048.
> 
> > echo 2048 > /sys/block/dm-1/queue/max_sectors_kb
> > dd if=/dev/zero of=/dev/mapper/thin bs=1M count=10K
> 10240+0 records in
> 10240+0 records out
> 10737418240 bytes (11 GB) copied, 223.517 s, 48.0 MB/s
> 
> The performance is nearly the same, and the block_dump message shows
> that the IO block_size is still 8 sectors per bio.
> To test if the direct IO does the trick, I try:
> 
> > dd if=/dev/zero of=/dev/mapper/thin oflag=direct bs=1M count=10K
> 10240+0 records in
> 10240+0 records out
> 10737418240 bytes (11 GB) copied, 192.099 s, 55.9 MB/s
> 
> However, the block_dump message shows the following line repeatedly:
> [614644.643377] dd(20404): WRITE block 942080 on dm-1 (1344 sectors)
> [614644.643398] dd(20404): WRITE block 943424 on dm-1 (704 sectors)
> 
> It looks like each IO request of dd has been split into 2 bios with
> 1344 and 704 sectors.
> In this circumstances, we can never follow the shorter path in dm-thin
> since a single BIO seldom overwrites the whole 1MB block.
> I also perform the same experiment with pool size equals to 512KB, and
> everything works as expected.
> 
> So here are my questions:
> 1. Is there anything else I can do to force or hint the kernel to
> submit 1MB size bio when it is possible? Or the only thing I can do is
> to stick with the block size lower or equal to 512KB instead?

I tried to reproduce but couldn't using a 3.12-rc1 kernel:

(the thin-pool is using a blocksize of 1024k, and the pool's underlying
data device has max_sectors_kb of 1024.. all layers of the dm devices
inherited that max_sectors_kb too just by the normal block layer's limit
stacking)

# dd if=/dev/zero of=/dev/vg/thinlv bs=1024k count=10 oflag=direct

dd(16494): WRITE block 0 on dm-4 (2048 sectors)
dd(16494): WRITE block 2048 on dm-4 (2048 sectors)
dd(16494): WRITE block 4096 on dm-4 (2048 sectors)
dd(16494): WRITE block 6144 on dm-4 (2048 sectors)
dd(16494): WRITE block 8192 on dm-4 (2048 sectors)
dd(16494): WRITE block 10240 on dm-4 (2048 sectors)
dd(16494): WRITE block 12288 on dm-4 (2048 sectors)
dd(16494): WRITE block 14336 on dm-4 (2048 sectors)
dd(16494): WRITE block 16384 on dm-4 (2048 sectors)
dd(16494): WRITE block 18432 on dm-4 (2048 sectors)

A dd that uses buffered IO will issue $PAGE_SIZE IOs.  So 4K IO in most
cases, unless the upper layers (or application) takes care to construct
larger IOs (like XFS does).

With XFS I see:

# dd if=/dev/zero of=/mnt/test bs=1024k count=1

dd(16708): WRITE block 96 on dm-4 (1952 sectors)
dd(16708): WRITE block 2048 on dm-4 (96 sectors)

# dd if=/dev/zero of=/mnt/test bs=1024k count=10

dd(16838): WRITE block 96 on dm-4 (1952 sectors)
dd(16838): WRITE block 2048 on dm-4 (96 sectors)
dd(16838): WRITE block 2144 on dm-4 (1952 sectors)
dd(16838): WRITE block 4096 on dm-4 (96 sectors)
dd(16838): WRITE block 4192 on dm-4 (1952 sectors)
dd(16838): WRITE block 6144 on dm-4 (96 sectors)
dd(16838): WRITE block 6240 on dm-4 (1952 sectors)
dd(16838): WRITE block 8192 on dm-4 (96 sectors)
dd(16838): WRITE block 8288 on dm-4 (1952 sectors)
dd(16838): WRITE block 10240 on dm-4 (96 sectors)
dd(16838): WRITE block 10336 on dm-4 (1952 sectors)
dd(16838): WRITE block 12288 on dm-4 (96 sectors)
dd(16838): WRITE block 12384 on dm-4 (1952 sectors)
dd(16838): WRITE block 14336 on dm-4 (96 sectors)
dd(16838): WRITE block 14432 on dm-4 (1952 sectors)
dd(16838): WRITE block 16384 on dm-4 (96 sectors)
dd(16838): WRITE block 16480 on dm-4 (1952 sectors)
dd(16838): WRITE block 18432 on dm-4 (96 sectors)
dd(16838): WRITE block 18528 on dm-4 (1952 sectors)
dd(16838): WRITE block 20480 on dm-4 (96 sectors)

The splitting of the IOs to not always be 2048 sectors is likely due to
the layout of XFS ontop of the thin LV.  Meaning the data area of XFS is
offset so as not to be perfectly aligned to the underlying thin LV's
blocksize.

Dave? Carlos?  Any hints on how I can prove this misalignment by
inspecting the XFS data areas relative to the underlying device?

> 2. Should the max_sectors_kb's attribute of the thin device be
> automatically set to block size?

max_sectors_kb is bound by max_hw_sectors_kb.  So max_sectors_kb may not
be able to scale to the thin-pool's blocksize.

But if max_sectors_kb can be set to the blocksize it isn't unreasonable
to do this.  I'll think a bit more about this.