[docs] Using bssplit with zonal disks uses min bssplit not sector size as min

[ckenna]

Please acknowledge the following before creating a ticket

• [✅] I have read the GitHub issues section of REPORTING-BUGS.

Description of the bug:

This is mostly a bug to consider clarifying some documentation. The short version of this is that when using a zonal device, the writes have to be aligned to physical block size. Couple that with using a bssplit setting, and I discovered that the min block size is coming from bssplit and not from the device itself.

So two possible outcomes are: Add this to fio docs and/or change the ZBD code to get min_bs from the zonal device and not bssplit.

Extra background:

For example, when I create a read job using bssplit=6ki/50:10ki/50 (full job file is below), I see only 6ki reads from fio.

I found a place in the zbd code where fio requires all writes to be aligned to the minimum block size, but it's taking the min block size from `bssplit. I traced the location down to here with some debug prints I added:

fio/zbd.c

Lines 2013 to 2023 in a191635

	/*
	* Make sure the I/O does not cross over the zone wp position.
	*/
	new_len = min((unsigned long long)io_u->buflen,
	(unsigned long long)(zb->wp - io_u->offset));
	new_len = new_len / min_bs * min_bs;
	if (new_len < io_u->buflen) {
	io_u->buflen = new_len;
	dprint(FD_IO, "Changed length from %u into %llu\n",
	orig_len, io_u->buflen);
	}

Here is the output of my debug printing (lines I added are prefixed with kenna:) and some commentary from me:

# On line 2018, the code unconditionally executes this:
# new_len = new_len / min_bs * min_bs
# https://github.com/axboe/fio/blob/a191635ad42893a60d0a917b27e97e3fc73f7fee/zbd.c#L2018
io       10439 kenna: after doing new_len / min_bs (6144) * min_bs => new_len = 6144
# So after doing 10240 / 6144 * 6144 in integer division, we are left with 6144.
# That's validated by the original debug printf.
io       10439 Changed length from 10240 into 6144
io       10439 kenna: because new_len (6144) < io_u->buflen (6144)

The minimum block size is 6144 but probably should be based on the actual disk block size.

Environment: Linux 5.10.210

fio version: fio-3.38-4-gcd56-dirty (I am assuming dirty is b/c I added my own dprintf statements. git is at cd56c0)

Reproduction steps

[global]
direct=1
time_based
runtime=5s
ioengine=libaio
filename=/dev/sda
zonemode=zbd
zonesize=268435456
ignore_zone_limits=1
thread=1
lat_percentiles=1

[read]
rw=randread
# Does NOT issue any 10ki reads, issues 6k reads, no 4k reads.
bssplit=6ki/50:10ki/50

write_iolog=/tmp/io-log.txt

[vincentkfu]

@kawasaki @damien-lemoal Do you have any thoughts about this?

[damien-lemoal]

Note: Zonal disk -> zoned block device (or zoned disk or SMR disk), please. "zonal" is not the correct term.

Regarding "The short version of this is that when using a zonal device, the writes have to be aligned to physical block size", yes, that is a hardware requirement. Not sure if we actually check that anywhere in fio/zbd.c, but we should. Will check.

Regarding the bssplit use, as far as I understand it, all IOs should always have an offset aligned to the block size used, unless bs_unaligned is also set. And that is not special to zbd.c. Would need to check that again. The code in zbd.c does ignore bs_unaligned though, so that probably needs to be fixed after checking what the regular use case (not zonemode=zbd) does as we should do the same.