►
From YouTube: July 2021 OpenZFS Leadership Meeting
Description
At this month's meeting we discussed: OpenZFS Conference; ASAN; changing default property values.
https://docs.google.com/document/d/1w2jv2XVYFmBVvG1EGf-9A5HBVsjAYoLIFZAnWHhV-BM/edit#
A
All
right
so
first
thing
on
today's
agenda
is
the
open,
zfs,
developer
summit,
dates
and
planning
the
summit.
If
you
look
for
a
go,
look
at
the
summit
home
page
on
the
opencfs.org
website,
these
key
dates
are
up
to
date
at
the
moment.
So
registration
is
opening
the
23rd.
So
in
a
few
days
please
submit
any
abstracts
proposals
for
presentations
if
you
have
them
so
that
we
can
get
some
content
going,
they
need
that
needs
to
be
done
before
the
end
of
august
august.
A
The
speakers
will
be
selected
selected
notified
by
september
10th,
and
the
conference
is
scheduled
for
the
8th
and
9th
of
november,
so
the
conference
is
currently
scheduled
to
be
a
remote
conference
again
this
year,
but
there
is,
I
know,
some
interest
in
the
possibility
of
maybe
doing
a
hybrid
conference
this
year,
where
some
people
might
show
up
and
yeah.
I
can
see
from
some
reactions
out
there.
A
Typically,
there
are
some
strong
opinions,
both
directions,
mostly.
What
I
want
to
get
a
feel
for
amongst
this
group
is:
if
there
was
an
in-person
option,
would
anybody
be
interested
in
showing
up
in
person
raise
your
hand.
B
A
All
right
yeah,
I
mean
you
know
the
issue
here.
I
think
for
most
of
us
and
for
me
as
well
is
that
you
know
I'm
sure
most
of
all
you
are
are
vaccinated,
but
with
the
variations
out
there
these
days
and
the
potential
sort
of
spike
in
cases
which
might
come
in
this
fall,
that's
being
predicted.
A
I
think
that
there's
there's
definitely
a
lot
of
unease
with
respect
to
that.
So
yeah,
I
don't
know,
what's
going
to
happen
exactly,
but
but
matt
wanted
to
get
a
feel
for
you
know.
If
there
was
a
lot
of
interest
in
in
some
sort
of
hybrid,
then
then
they
might
do
something
about
that,
but
obviously
there's
a
lot
of
logistics
that
occurs
in
terms
of
a
venue
and
that
kind
of
thing.
A
A
A
Right
next
item
on
the
agenda
is
a-san
running
z-test
with
the
address
sanitizer.
So
I
think
in
general
this
is
a
good
idea.
The
the
address
sanitizer
malik
debuggers
would
add
some
nice
additional
checking
for
us
during
z.
Test
runs
the
the
reason
this
hasn't
been
done
in
the
past.
I
don't
know
all
the
details.
I
know
that
matt
when
he
started
playing
with
it.
He
ran
into
some
issues
with
the
what
was
it
the
zfs
standard
memalic
some.
A
Yeah
reports
you
use
after
free
with
zstd
mempool
code,
so
he
created
an
issue
one
two
one.
Let's
say
one,
two,
two
one:
five
for
those
interested
in
checking
it
out.
It.
A
He
was
unable
sort
of
to
get
a
good
run
without
running
into
his
shoes
here,
so
it
would
just
creep.
Crapping
out
in
the
strategy,
ended
up
sort
of
commenting
out
the
code,
basically
and
bypassing
it
in
order
to
get
things
to
run
cleanly
for
him.
So
that's
kind
of
an
issue.
It
needs
to
be
tracked
down
and
figured
out.
It's
doesn't
seem
to
be
something
which
necessarily
reproduces
for
everyone.
So
take
a
look
at
the
issue.
There's
people
who
said
one's
fine
for
me.
A
I
think,
though,
that
we
need
to
figure
out
how
to
get
the
code
fixed
if
necessary
or
or
at
least
work
around
this
issue,
so
that
we
can
get
the
azure
sanitizer
running
regularly
to
do
regular,
builds
with
azure
sanitizer
enabled
and
get
our
z
test
runs.
So
we
can
cover
memory
issues
that
might
might
be
latent
in
the
code
or
be
introduced
by
changes
that
are
coming
in
yeah.
C
Mark
johnston
from
the
freebsd
project
also
has
a
pull
request
open
to
fix
some
issues.
He
found
with
the
kernel
memory
sanitizer
while
actually
running
zfs
in
the
kernel
in
freebsd
yeah.
A
So
I
think
anything
like
that
is
is
a
win
for
us
yeah,
the
data,
and
it
picks
it
in
like
that.
I
ran
a
bunch
of
runs
with
the
edge
of
sanitizer
a
week
or
two
ago,
because
we
were
there.
There
was
a
few
issues
which
we
encountered
a
delphics
around
memory,
issues
that
I
couldn't
figure
out
exactly
what
was
going
on.
A
Their
sanitizer
was
able
to
to
find
one
issue
for
me,
which
was
an
issue
which
I'd
sort
of
already
partially
identified,
but
I
think
just
having
that
extra
tool
in
our
pockets
is
really
handy
for
these
kind
of
things,
because
memory,
corruption,
issues
when
you
encounter
them
are
a
to
track
down,
sometimes
like
try
to
access
freed
memories
like
how
did
that
happen.
Try
to
access.
A
You
know
whatever
it's
like
what
happened
here,
and
so
without
that
sanitizer
or
something
like
a
tool
in
there
to
track
it
down
the
point
where
it
actually
occurs.
It's
it's
really
really
difficult
to
get
that
kind
of
that
stuff
tracked
down.
So
anybody
have
any
any
other
comments
around
leveraging
address
and
hazards
or
some
objections
to
it.
A
C
A
You
know
it's
kind
of
a
z-test
issue
in
general.
I
think
you
know
right
now.
Z-Test
is
pretty
good.
It's
it's
runs
pretty
clean.
There
are
a
few
kind
of
known
issues
around
z-test,
sometimes
a
crop
up,
but
it's
hard.
A
You
know
I
I'm
now
in
my
role
of
looking
at
a
lot
of
the
pr's
that
are
coming
in
and
looking
at
the
test
results
and
trying
to
determine
whether
or
not
there's
anything
that
you
followed
up
on
and,
and
it
can
be
quite
difficult
and
even
you
know
z-test,
because
it's
not
entirely
reproducible
you'll
get
a
z-test
failure
which
is
like
well,
that's
not
in
the
code
very
likely
to
be
in
the
code.
That's
that's
being
you
know
being
requested
to
be
merged
right
now.
A
This
is
probably
something
that
someone
else
introduced
and
and
trying
to
sort
of
trace
that
back
and
and
even
getting
that
person
who
might
have
done
that
pr
to
work
on
it
can
be
can
be
tricky.
So
it's
it's
nice
when
you
can
discover
these
things
at
the
point
of
integration,
attempt
and
say
hey,
no,
it's
not
quite
ready.
Yet
you
should.
A
What
happened
here
in
your
bits
and
also
it
gives
you
a
context
to
work
in
which
is
a
lot
easier
when
you
have
the
entire
code
business.
So
just
I
made
these
five
changes
so
something
in
here
may
have
been
causing
this
problem.
So
I
think
all
that
is
is
important.
Stuff.
C
Yeah
there's
sometimes
where
a
test
will
fail
and
it'll
it'll
say
why
it
failed,
but
it
won't
say
what
led
up
to
it
like.
There
seems
to
be
missing
context
or
like
the
command
that
I
like,
I
think,
if
you
have
g-test
underscore,
must
or
whatever
for
something
that
has
to
return
success.
C
A
C
The
data
set
doesn't
exist,
it's
like
it,
just
the
cache,
file's
been
updated
or
it
didn't
sync
yet
or
or
something,
but
it's
like
it
doesn't
actually
say
it.
You
know
it
gives
me
standard
error
from
the
command,
but
it
doesn't
tell
me
what
the
command
it
tried
to
run
was.
So
I
I
have
trouble
trying
to
recreate
this
exactly.
I
don't
need
enough
context,
sometimes
to
feel
what's
happened.
There.
A
Yeah,
I
know
I've.
I've
felt
I've
had
this
animation,
sometimes
dependency
test
problems,
all
right
or
or
zts
problems.
For
that
matter.
Let's
see.
A
A
Introducing
or
trying
to
introduce
the
outer
sanitizer
for
us
as
part
of
our
z-test.
I
know
brian
can't
be
here
today:
either
he's
he's
not
available,
and
he
was
also
you
know
in
favor
of
this.
So
so
I
think
there's
definitely
needs
to
be
a
push
here
to
see
we
can.
We
can
do
that
to
to
get
it
introduced
as
part
of
our
standard
test
runs
for
integration
prior.
A
Animation,
someone
asked
for
a
status
update
on
the
panzora
temple
dedupe.
I
matt
says
that
looks
like
the
latest
update
was
in
april
of
2020
and
they
have
deep
prioritized
this
upstream
work,
so
he
didn't
seem
like
he
thought
it
was
very.
A
It
didn't
seem
like
it
was
an
active
project
at
the
moment.
I
don't
know,
I
don't
have
any
other
information
beyond
that.
Does
anybody
else
have
any
any
knowledge
or
any
information
about
that.
C
A
Yeah
has
anybody
had
any
experience
with
with
their
temple
video
project.
C
A
Yeah,
I
I
weighed
way
years
ago
and
zurich
came
and
talked
to
us
that
I
think
it
was
oracle.
A
Their
technology,
it
was
interesting,
there's
interesting,
stuff
they're.
Definitely
I
have
not
followed
it
at
all
since
then,
so
don't
know
much.
B
I
assume
there's
no
one
from
the
company
here:
yeah
yeah,
I
did
not
get
a
terribly
great
long-term
survival
rate
glim
from
from
the
last
time
I
spoke
with
them.
Okay,.
D
All
right,
can
you
hear
me
so
the
problem
is
that
extended
attributes
are
named
differently
on
different
platforms,
so
loomis
doesn't
have
namespaces.
So
all
the
names
are
just
raw
freebsd
has
a
couple
different
name
spaces
and
we
prefix
the
system
name
space
with
a
freebsd
colon
system,
colon
string
and
the
username
space
names
are
just
raw
like
on
a
lumos.
B
D
And
then
you
can
turn
that
off.
If
you
know
that
you
don't
care
and
then
there's
a
x,
add
or
compat
property
that
can
be
set
to
all
to
write
the
username
space
attributes
the
same
across
all
platforms
and
so
on
linux.
It
would
write
it
without
the
prefix
and
it
also
has
a
exciter
compat
equals
linux.
If
you
want
to
set
it
back
to
be
compatible
with
what
linux
currently
does.
D
The
one
kind
of
gotcha
with
this
is
that
it's
kind
of
an
on
disk
format,
change
for
linux,
and
so
I
added
a
cool
feature
that
activates,
when
you
write
an
extended
attribute
in
the
new
format
on
linux,
and
I'm
mostly
looking
for
a
design,
review
and
code
review
on
this
at
this
point,
we're
starting
to
ship
it
for
trunas
and
so
in
beta
builds
and
so
we'd
like
to
get
it
upstream
asap
and
then
go
through
any
design
iteration.
That
needs
to
happen
before
we
get
too
far.
Along
with
that.
D
The
the
main
thing
I
have
questions
about
is
the
fallback
option
has
kind
of
a
performance
penalty,
because
you
have
to
do
a
lookup
again
if
you
fail
the
first
time
and
so
it'd
be
nice
to
be
able
to
default,
to
offer
that
and
then
automatically
enable
it
when
it
needs
to
be,
but
I'm
still
kind
of
investigating
how
to
actually
mechanically
do
that
in
cfs.
A
D
B
Okay,
does
linux
have
system
properties
that
do
not
have
a
prefix.
D
B
A
D
Just
can't
get
it
a
slew
of
attributes
and
kind
of
a
possibly
work
like
a
roundabout
security
issue.
Is
that
on
other
platforms,
if
you
can
write
in
the
username
space
with
a
name
that
looks
like
the
prefix
of
like
a
system
name
space
on
linux
and
then
import
the
pool
on
another
platform,
then
you
can
do
stuff.
You
probably
shouldn't
be
able
to
do,
and
so
I
added
checks
that
deny
using
those
prefixes.
A
D
Yeah
I've
had
the
pull
requests
open
for
comments
for
a
couple
months
and
haven't
got
much
feedback
on
it.
Yet
so,
okay.
E
Sorry
about
that
ryan,
alexander,
provided
some
feedback,
I
think
brian
will
probably
be
the
other
person
that
can
give
more
design
kind
of
review
our.
I
will
ping
him
once.
A
E
E
You
know
I
do
not
see
any
comments
from
brian
there,
yet
all
right.
A
A
That,
I
think,
is
all
we
have
on
our
agenda
formal
agenda
today.
Is
there
any
topics
that
anyone.
C
Would
like
to
remember,
I
threw
a
couple
things
on
the
end
there,
the
linux
namespace
support
proquest.
F
C
Is
ready
now?
So
if
you
want
to
look
at
that,
we've
been
testing
it
and
it's
seems
to
be
working.
C
We
had
to
revert
one
change
that
was
committed
quite
a
while
back
for
I
think
linux,
4.9
compatibility
or
something
it
changed.
The
way
you
looked
up
the
user
id
so
that,
if
you
were
in
a
namespace,
it
tried
to
look
up
the
id
relative
to
the
namespace,
but
from
zfs
perspective
on
disk,
we
always
want
to
store
the
user
id
relative
to
the
root
name
space.
C
If
you
tried
to
see
h
on
a
file
inside
a
namespace,
so
the
the
first
commit
of
that
pull
request
is
actually
to
revert
some
of
the
bits
of
that
that
were
causing
the
problem
and
then
the
rest
introduces
the
zfs
user
ns
command,
which
allows
you
to
delegate
a
data
set
to
a
username
space
on
linux,
similar
to
how
you
delegate
it
to
a
zone
or
jl1,
solaris
or
bsd,
and.
E
C
In
if
you're,
when
you're
in
the
namespace,
you
can
only
see
the
datasets
you're
allowed
to
see,
and
it
has
all
the
right
vfs
flags
for
linux
to
be
able
to
mount
as
an
unprivileged
user
inside
that
namespace
and
so
on,
and
so
it
provides
about
the
same
feature
set
and
I
added
updated
photographs
ryan
found
a
in
the
user
space
version
of
live
spl
or
whatever
it
was
breaking
the
freebsd
tests.
So
I
added
an
iftaft
to
fix
that.
C
We
do
the
right
thing
in
the
actual
spl
kernel
module
stuff,
but
the
the
use
face
version
for
the
tests
was
always
setting
the
global
zone
id
number
to
the
linux
one,
whereas
on
illumios
and
freebsd
the
global
zone
is
zero,
not
right.
Some
random
large
number
that
linux
picks
right
right.
E
C
To
do
that,
no
I've
not
had
a
chance
to
look
into
that
one.
Yet
the
stick
style
thing
is
the
check.
Api
is
failing
I'll.
D
C
A
C
D
A
pr
that
generates
those
in
the
ci
job
now,
if
it
fails,
so
you
can
just
download
the
new
abi
files
from
the
ci.
I
don't
think
it's
been
merged
yet,
but
okay.
G
E
So
maybe
I'll
I'll
try
to
get
that
one
more
and
then
I'll
and
you
you
can
take
advantage
of
that
awesome.
C
A
C
That
one
I
I
got
I've
I
went
through
and
double
checked
and
cleaned
up
all
the
xxx
stuff.
So
that's
all
sorted
out
and
I'm
working
with
dylan
cochran
to
write
some
tests
for
it.
D
C
It
doesn't
have
enough
tests,
for
you
know,
make
sure
you
can
set
a
property,
read
it
back
and
then
unset
it
and
things
like
that
and
the
other
one
was
for
fixing
the
name
of
the
v
dev
like.
If
you
want
the
video
in
the
cache
file
to
reference
a
different
device
name
than
it
does
right
now
you
can
set
it,
but
we
want
to
probably
put
some
tests
in
there
to
make
sure
it
fails
when
it
should
and
succeeds
when
it
should.
A
A
Yeah
I'd
like
to
see:
let's
get
that
one
wrapped
up
so
yeah
does
anything
you
need
me
for
there
allen.
Let
me
know,
depending
on
that,
I
think
the
the
alec
stuff
is
solid
in
there
at
this
point,
but
you
may
I
mean
it's
sort
of
tested
by
the
device
removal
path
now,
because
that's
the
way
the
device
removal
works,
but
but
yeah
there's
other
uses
for
no
out.
A
We
should
probably
have
at
least
a
test
case
or
two
around
verifying
the
note
and
said
you
know,
alex
with
respect
to
space
usage,
etc.
One
of
the
because
I
think
we
talked
about
this
in
my
private
previous
meeting-
was
setting
no
alec
property
on
a
v.
Dev
removes
the
more
space
than
you
might
expect
from
the
available
capacity,
because
it
removes
the
storage
capacity
of
the
drive
itself
saying
all
right.
We
have.
A
The
storage
obviously
is
gone
from
your
ability
to
space,
but
also
removes
the
space
that's
using
currently
from
the
storage
capacity,
because
if
you
were
to
overwrite
that
data
you'll
end
up
being
rewritten
into
the
rest
of
the
data,
so
it
would
look
as
though,
like
you
had
not
actually
written
your
data
but
somehow
used
up
more
space
in
the
pool.
So
you
have
to
be
able
to
count
that
space
as
well.
So.
A
No,
no
so,
for
example,
if
you
were
to
say
you
had
a
pool
which
had
simplest
case
actually,
three
three
drives
of
10
terabytes
with
30
terabytes
of
space,
you
know
remove,
see.
I
want
to
remove
one
of
them.
That's
all
right
and
you
have
them
all
half
full.
You
used
up
15
terabytes
of
space
within
your
30
terabyte
space.
This
is
a
little
bit
fuzzy,
because
there's
actually
actually
considered
like
this.
A
When
you
remove
that
as
to
remove
that
that
10
terabyte
drive
says
okay,
I'm
going
to
tell
you
that
out
of
your
of
total
30,
I'm
removing
10
available
plus
another
five
available,
so
that
you
really
only
have
you
know
the
available
space
now
is,
is
only
going
to
be
five
terabytes
in
your
resulting
pool,
because
that's
really
all
that's
that's
absolutely
available
for
new
data.
A
A
So
it's
a
little
bit
fuzzy
because
like
well,
you
know,
if
you
don't
ever
rewrite
that
data.
That's
in
that
no
out
part
of
the
pool.
In
theory,
you
have
10
terabytes
of
space.
You
could
fill
up.
The
benefit
were
to
attempt
to
if
you
were
to
fill
that
up
and
then
attempt
to
rewrite
some
of
your
data,
which
was
in
the
no
alec
part
of
the
pool
which
you're
allowed
to
do.
But
no
it
doesn't
mean
you
can't.
A
And
simply
modify
it
and
say
I
open
the
space.
I
can't
do
that
because
you're
used
to
call
space.
So
it's
it's
a
little
bit
it's
a
little
bit
funny,
but
it
does
follow
sort
of
our
our
space
allocation
model
reasonably
accurately.
It's
just.
You
have
to
think
about
a
little
bit
just
to
understand.
A
C
A
C
And
then
the
other
thing
I
was
gonna
talk
about
was
we've
started
a
project
to
improve
the
z-vol
performance
on
linux
or
g-vol
in
general,
but
specifically
on
linux,
comparing
it
with
a
block
size
of
16k,
we're
seeing
about
a
12
slower
on
ubuntu
than
on
freebsd,
using
the
same
checkout
of
of
open
zfs
and
that's
in
gm
modem,
freebsd
after
full
mode
or
whatever.
They
call
it
now,
and
I
think
in
plain
dev
mode
it
might
actually
be
even
a
bigger
difference.
C
So
we're
just
getting
started
with
that.
But
if
anybody
is
having
especially
pathological
workloads
or
something
who
knows
how
to
to
make
a
z-vol
really
fall
on
its
knees,
we'd
love
to
see
what
some
of
those
are,
so
that
we
can
target
some
of
those
spots
to
gain
to
them.
A
So,
several
years
ago,
I
did
some
work
trying
to
understand
why
the
z-volt
performance
mp,
was
not
that
good
on
on
links,
because
I
was
running
into
submissions
there
back
when
I
was
working
in
craig
and
my
recollection,
I
don't
remember
all
the
details
was
that
there's
some
some
pretty
clunky
use
of
the
block
layer
in
by
the
z-ball
code
in
zfs,
for
for
in
linux,
where
they
were,
there
were
ending
up
doing
a
lot
of
things.
A
You
know
a
a
sector
at
a
time
in
some
of
these
interface,
some
lower
level
layers
of
the
block
interfacing
with
the
block
layers.
It
was
like.
Why
aren't
you
doing
this
in
larger
chunks,
sort
of
thing
with
my
recollection
of
dishes?
I've
seen,
but
I
don't.
A
I
just
don't
remember
details
though
I
don't
know
if
that
that
helps
you
at
all.
But
that's
that
was
my
experience.
I
I
also,
I
definitely
had
the
experience
saying
this
is
something
is
wrong
here
with.
C
Yeah,
I
think,
before
a
commit
from
matt
from
a
while
ago,
where
he
added
a
task
you
and
so
on
that
was
kind
of
direct
dispatch,
and
so
it
was
really
bad
you.
You
really
noticed
that
I
was
doing
one
sector
at
a
time.
It's
it's
a
bit
better
now,
but
having
that
one
task,
you
where
it
might
make.
D
C
Sense
to
use
a
more
native
interface
and
let
the
linux
cueing
layer
do
it
or
something
right
right,
looking
at
a
bunch
of
different
things
there,
but
if
anybody
knows
something
that
really
performs
poorly
compared
to
just
you
know,
fio
or
the
postgres
benchmark
or
something
you
know
we're
going
to
try
to
fix
it.
So
you'll
get
your
fix
for
free.
If
you,
let
me
know
what
the
problem
is.
C
And
the
last
one
is
something
that
came
up
a
long
time
ago
and
we
talked
about,
I
think,
everybody's
in
agreement,
but
I
don't
think
anybody
ever
did
it
was
we
wanted
to
change
the
default
when
you
create
a
new
pool
to
have
lz4
compression
on
by
default.
Since
there's
almost
no
reason,
you
wouldn't
want
it,
and
if
you
do,
you
would
know
that
you
don't
want
it
and
could
turn
it
off,
whereas
most
people,
I
think,
would
want
the
you
know.
C
Lz4
is
almost
free
and
has
there's
also
just
so
much
code
in
zfs.
That's
if
compression
is
not
off
than
do
this,
like
the
zero
suppression
and
stuff
that
it
kind
of
feels
like
having
compressed
default
on,
would
be
a
big
win
and
a
small
change
and.
D
It
sounds
like
a
good
idea
to
me.
It
makes
me
think
of
another
default
that
maybe
could
change
now.
Is
the
x
adder
equals
sa
before
freebsd
didn't
support
that
and
so
for
compatibility.
The
default
is
dur,
but
now
I've
implemented
it
for
freebsd.
So
we
could
change
that
if
you
wanted
to
too
that'd
be
a
little
performance
boost
too.
C
A
Yeah,
I
think,
there's
a
lot
of
history
there
around
compression,
whether
it
was
you
know
acceptable
to
have
a
compression
level
by
default.
We
certainly
had
discussion
at
various
times
around
whether
or
not
we
would
or
wouldn't
have
a
default
compression
in
there.
I.
D
A
I
think
there
was
some
concern
in
those
days
also
with
performance
and-
and
you
know,
almost
free
wasn't
quite
as
free
as
it
had
done
process
just
today,
as
it
was
back
then,
which
also
made
it
a
fact.
I
don't
remember
the
details.
G
A
H
H
We
find
eight
k's
just
to
switch
any
sufficient,
sequential
boundaries
throughput
even
16k
limits
are
put
quite
significantly,
but
8k
is
just
unacceptable,
in
my
opinion,
in
many
ways,
and
also
in
context
of
physical
block
size
of
disks
being
4k
now
quite
often
even
on
ssds
8k
wall
block
size
just
doesn't
allow
any
significant
compression
it's
practically
on
or
off
so
16k
blocks
also
should
compress
a
lot
better.
H
A
H
8K
is
just
an
acceptable.
16
case
is
bad
too,
very
bad,
but
at
least
twice
better
for
compression
8k
is
bad
and
also
on
many
tests.
I
regularly
see
that
our
presence
in
thread
can't
handle
more
than
quarter
million
blocks
per
second,
so
quarter
million
at
eight
k
gives
two
giga.
H
A
Yeah
yeah
kind
of
feeds
into
your
your
z-ball
performance
improvements.
There
alan.
C
Yeah
and
the
compression
thing,
because
yeah.
A
A
C
C
C
The
rights
together,
because
that
provides
much
better
performance
on
spinning
disks
and
we
you
can
turn
that
off
for
ssds.
But
I
wonder
if
we
actually
want
an
option
to
do
the
opposite
and
purposely
write
to
a
bunch
of
different
places
on
the
ssd
at
once,
to
activate
more
of
the
flash
cells
at
once
and
get
more
throughput.
C
Like
we've
seen
a
case
with
one
customer's
ssds,
where
you
know,
if
you're
writing
sequentially,
you
don't
get
nearly
the
same
speed
as
if
you
write
sequentially
to
these
very
different
offsets
and
if,
if
it
would
make
sense
to
purposely
grab
meta
slabs
from
very
different
parts
of
the
disk
and
write
to
them.
Concurrently,.
H
I
was
gonna
say
that
we
already
have
multiple
locators
now,
each
of
which
should
have
separate
meta
slabs
open.
So
if
you
write
several
files
same
time,
you
should
write
into
several
areas
of
ssd,
just
as
you
told,
unfortunately,
it
doesn't
work
for
several
z-volts.
At
this
point,
science,
we
see
only
one
z
wall
at
a
time,
but
that's
another
thing.
I'd
really
like
to
be
fixed.
H
H
H
Honestly,
I
don't
understand
I,
I
really
doubt
I
really
doubt
about
it:
the
slabs
already
pretty
big.
We
don't
have
so
many
of
them,
that's
from
one
side
from
another
site.
You
mentioned
lbi
bios
and
they
have
significantly
strong
feelings.
That
last
time
I
looked
there,
it
was
not
effect
effectively
used
for
present
science
science.
Some
pool
feature
I
forgot,
which
one
which
was
added
few
years
back
and
enabling
it
practically
disables
lba
bios.
We
shouldn't
have
it
this
day.
A
H
A
C
A
Right
and
I
might
worry
whether
or
not
you
were
impacting
the
the
durability
of
the
drives
in
some
fashion
by
sliding
across
a
larger
range
of
lbas,.
C
D
A
All
right
anything
else.
A
Everybody
have
a
presentation
ready
to
go
for
the
the
opens
your
best
match
coming
in.
C
G
D
D
C
I
they
didn't
show
up
in
the
list
for
the
the
push
I
did
before
the
meeting.
A
Yeah,
so
maybe
there's
something
wrong
with
the
test
framework
right
now.
I
know
there
was
some
issues
with
them
a
few
weeks
ago,
and
so
maybe
that
that
brian
said
that
they
need
some,
some
fixing
me.
So
maybe
that's
happening
or
you
die.
C
A
So
I
think
I
mentioned
I'm.
I
I
now
tony
and
I
and
john
kennedy
are
doing
a
lot
of
the
the
pr
management
now
in
the
in
the
gate.
So
if,
if
you
have
prs
that
are
languishing,
please
feel
free
to
ping
us
we
try
to
keep
on
top
of
things
by
the
line.
Particular
sometimes
let
things
slide
a
little
bit
until
alexander
pings
me
and
says
you
could
not
do
this
so
try
to
keep
I'll
try.
A
We
try
to
keep
up,
but
it
helps
to
get
a
thing
periodically.
If
you
say
hey,
what's
going
on
guys,
do
you
find
someone
to
do
this
or
just
ready
to
go
in
or
what
okay.
I
Think,
well,
I
think
matt
kind
of
suggested.
I
mentioned
one
of
my
prs
that
we're
doing
as
we've
been
pushing
the
apple
commits
or
prs
for
apple
support.
Yes,.
I
And
trying
to
doing
as
small
as
possible,
so
you
know,
makes
it
easier
for
everyone
to
do
code,
reviews
and
whatnot.
Most
of
them
are
real,
but
the
one
of
the
ones
we
have
at
the
moment
is
enabling
mounting
and
unmounting
of
snapshots
using
the
zfs
command
so
that
you
right
now
you
can
say
zfs
mount
and
then
data
set
so
we've
added
it.
So
you
can
say
cfs
mount
and
then
the
snapshot
and
the
reason
for
that
is
on
the
apple
kernel.
We
cannot
trigger
mounts
from
the
kernel.
I
It
just
can't
be
done.
So
it's
the
first
step.
We
felt
it
was
sort
of
easy
for
the
users
to
use
cfs
mount
and
then
snapshot,
and
it's
pretty
easy
to
understand.
I
suppose
it
doesn't
let
you
specify
where
or
anything
it
still
uses
the
same
dot,
zfs
directory.
Oh
interesting,.
A
I
Right,
it's
just
so
because
it
has
to
come
from
new
zealand.
You
have
to
trigger
it
from
new
zealand
and
in
a
future
pr
to
the
bitter
spoiler.
We
do
add
snapshot
mounting
events
to
zed
the
zfs
event
daemon,
so
that
we
can
issue
mount
requests
from
the
kernel
via
zed
zfs
mount.
I
I
Well,
the
question
is
at
the
moment
that
is
all
hash
if
apple,
since
it
changes
the
semantics
of
the
zfs
command.
I
That's
really
a
good
idea.
Yeah,
maybe.
C
C
I
I
Thing
and.
I
Could
be
a
temporary
mom
point
would
be
nice,
but
in
the
case
of
snapshots
it
you
know
it
goes
in
the
dot
zfs
directory
yeah.
That's
all
it
does,
but
yeah.
I
But
either
way
that's
kind
of
beyond
the
scope
of
that
right.
Yeah,
yeah,
you're.
B
Good
yep
talk
to
me
in
games
and
slack
the
proper
way
to
do
that
with
mac.
Os
is
with
discarb
no
of
them.
I'm
aware
of
that.
So,
but.
B
I
In
the
kernel
right,
you
trigger
something
that
makes
easy
land
wake
up
and
mountain
what
you
can
do
in
the
cur.
What
you
can
do
in
the
kernel
is,
you
can
make
a
device
appear
correct,
yes,
that
that
is
how
apfs
does
it.
You
know,
we've
we've
gone
down
the
road
we
also
make
suited
devices
and
so
on.
I
We
were
aware
of
all
of
that
and
in
the
end,
about
what
10
years
ago,
when
we
did
the
port,
we
felt
it
was
easiest
to
let
your
users
issue
the
mount
command
if
they
wanted
it
to
at
that
time
now
we
could
do
by
creating
pseudo-devices
and
getting
this
carb
involved,
and
so
on.
That
is
true,
but
it
doesn't
change
the
fact
that
it
has
to
come
from
user
land.
This
carb
has
to
execute
your
tool,
which
comes
from
the
file
system
thing,
so
you're
still
editing
commands.
If
you
will.
C
C
I
think
there's
more
places
where
I'd
like
us
to
use
the
the
snap
spec
thing
to
specify
a
range
of
snapshots
in
the
commands,
but
you
know
sometimes
I
think
I'd
just
be
a
convenient
way
shorthand
for
incremental
replication.
Just
this
snapshot
through
that
snapshot.
C
I
Yeah
I
mean
similar
to
that.
There's
we
add
a
bunch
of
underscore
os
functions.
A
good
example
for
that
is
with
rollback.
After
you
call
lib
zfs
to
do
rollback.
We
need
to
do
rollback
underscore
os
section
for
apple,
where
we
kick
finder
into
refreshing.
Otherwise
it
keeps
showing
the
old
information
and
that
could
all
be
in
hash
if
theft,
but
I
think
it's
nicer
to
just
have
empty
functions
on
the
platforms
that
don't
need
it,
and
we
already
do
that
for
zvo
weight.
C
F
I
did
have
a
question
or
thing
just
to
get
some
ideas
here
is
one
of
the
things
we
ran
into
on
a
lumos
here
recently,
or
we
discovered
was
because
of
the
kind
of
the
ordering
of
porting
of
stuff
the
the
encryption
features.
We
never
got
the
fix
that
included
the
project
d
node
as
part
of
the
hash
which
got
fixed
in
between
releases,
which
technically
was
a
on
format.
Disk
change
for
encrypted
data
sets,
but
because
it
was
between
releases,
it
was
kind
of
well.
F
F
Unfortunately,
for
a
lumos,
though
it
is
out
in
the
wild
and
so
the
question
in
terms
of
the
best
way
to
kind
of
upgrade
upgrade
because
we'd
like
to
you
know,
you
know
the
goal
is
eventually
to
try
to
get
everything
combined
at
some
point
in
the
future.
F
It's
just
a
matter
of
doing
all
the
work,
and
so
there
are
people
that
do
like
to
be
able
to
exchange
pools
between
the
platforms,
which
right
now
is
an
issue,
and
so
you
know
the
issue
so
tr
and
thinking
in
terms
of
how
to
fix
it.
You
know
we
can.
Of
course
we
can
add
in
the
fix
that
you
know,
adds
the
project
d
node
to
the
hash
as
well
as
even
I
have,
and
I
just
I
tested
it,
the
other
night.
F
You
know
we're
okay,
if
it,
if
it
fails,
we
can
try
it
with
just
the
two
you
know
from
like
backwards
compatible,
but
then
okay.
Well,
if
you
still
write
out
the
new
one
with
all
three,
then
of
course
then,
then
immediately
then
your
pool
becomes.
F
C
There
until
you
recreated
the
data
set
and
or
whatever,
and
so
like
you're,
saying
kind
of,
like
a
the
per
data
set
feature
flag
thing
like
we
have,
for
you
know,
zed
standard
and
so
on.
It's
not
actually
considered
active.
Unless
you
have
data
sets
that
are
using
it
and
then
once
you
destroy
the
last
data
set
that
was
using
it,
it
reverts
to
being
enabled
as
a
way
to
tell
whether
you
need
to
worry
about
the
backwards
shim
or
not.
F
Yeah,
although
I
think
the
errata
is
like
per
pool,
because
I
think
it's
actually
a
feature
of
the
spa,
okay,
which
again
it's
like
okay-
well,
it's
really,
you
know
technically
per
data
set
and
so.
C
Yeah
the
feature
flags
like
the
way
we
did
it
for
z,
standard
and
and
just
things
like
sha,
512
and
so
on,
maybe
makes
the
most
sense
for
there
do.
We
have
the
concept
of
a
feature
where
you
don't
want
it
on
usually.
F
Or
the
other
thing
is
what
I'm
going
back.
We
are.
The
other
thing
was
to
create
like
a
dummy
feature.
Just
automatically
like
on
open
zfs
would
just
be
enabled
basically
on
any
pool,
but
it's
like
does
nothing
versus
on
a
lumos.
Okay.
If
we
see
that
flag,
okay,
it
means,
then
we,
you
know
we
use
all
three.
Otherwise
we
use
just
the
two,
at
least
when
writing,
although
maybe
you
know,
there's
still
some
debate,
you
know,
would
we
want?
F
You
know
you
know
if
we'd
want
to
be
able
to,
you
know,
control
the
backwards,
compatibility
or
not
just
because
technically
you
know
obviously
not
having
the
project
denode
in
the
hash.
Could
you
have
some
potential
implications
and
you
maybe
some
people
may
want
you
know?
Well,
no!
No.
F
I
want
that,
although
I
also,
although
there's
a
good
chance
that
there's
probably
very
few
if
anyone
using
the
project
features
on
a
lumos
but-
and
so
I
just
you
know
in
terms
of
anyone,
had
any
suggestions,
so
we
can
do
that
because,
like
I
said
I
can
just
you
know
basically
transparently
without
you
know
the
the
operator
ever
knowing
I
can.
I
can
fix
the
data
set
and
of
course,
once
you
do
that
you
can,
you
can
never
go
back
on
alumos,
which
is
you
know
the
issue.
F
I
could
you
know
we
could
do
tunables,
you
know,
so
the
districts
could
control
that
just
unfortunately,
because
I
think
the
project
feature
got
ported
after
encryption,
which
is
how
I
think
it
got
la
that
bit
got
lost
basically.
I
I
After
that,
oh
on
linux,
it
was
done
first,
so
we
have
the
same
problem
and
we,
as
you
suggested
we
test
two
first
and
if
that
fails,
we
then
tested
the
ashes
and
then
upgrade
and
upgrade
for
the
user.
And,
yes,
they
can't
go
back,
but
we
consider
that
to
be
like
zpool
upgrade.
You
just
can't
go
back,
but
either
way
whatever
solution
you
find
we'd
like
to
be
part
of
that
conversation,
because
we
also
want
to
have
a
better
solution.
F
And
so
like
I
said
so,
the
thought
was
maybe
doing
something
with
the
feature,
but
then,
like
I
said
it
would
require.
You
know
something
on.
You
know
zfs
on
linux,
even
though
maybe
like
on
zfs
on
linux.
It
may
just
be
like
a
dummy
feature
that
does
nothing
anywhere
else,
but
you
know
kind
of
the
question
I
don't
know.
If
there
would.
You
know
people
be
okay
with
that
or
is
that
gonna
cause?
You
know.
A
Concern
the
adjacent
idea
would
be
that
on
linux
you
would
suddenly
there
would
appear
this
new
feature,
which
is
represents
the
the
three
or
hatch
base.
But
it's
always
there.
It's
like
okay,
yeah,
it's
like
we
forgot,
just,
never
saw
the
feature
as
as
being
a
bit.
You
know
visible
before,
but
now
suddenly
it
shows
up
as
being
there
on
every
linux,
because
it's
it's
drawn
everywhere.
C
Because
yeah
I
was
thinking
is
like:
would
you
want
a
feature
flag
that
is
only
active
on
the
data
sets
that
are
not
fixed
or
something
to
that
effect
so
that
it
properly
like,
if
you've
only
been
doing
it
with
the
two
hashes
and
you
import
it
on
a
version
that
does
all
three?
Does
it
cause
problems
or
is
it
just
a
problem
if
you
wrote
with
a
newer
one
and
then
try
to
read
it
with
an
older
one
or
I
guess,
if
you
have
a
data
set,
that's
that
has
this
bug.
F
C
If
you
made
it
a
feature
flag
that
only
existed
on
the
platforms
that
had
the
problem
then-
and
it
was
activated
where
the
problem
is
present-
then
it
would
block
it
importing
on
systems
that
don't
have
the
problem,
but
as
long
as
the
but
the
flag,
the
feature
flag
is
there,
but
it's
just
enabled
because
there
are
no
damage
data
sets.
Then
it's
considered
read-only
compatible
or
even
rewrite
compatible.
I
guess
like
we
don't
have
that
concept.
Do
we
so.
D
F
And
that's
what
I
thought
was
actually
kind
of
the
opposite.
We
have
like
a
feature
where,
like
I
said
on,
you
know
like
on
you
know,
on
open
zfs.
It
really
you
know
when
it's
there.
It
really
does
nothing.
You
know,
as
I
say
it's
kind
of
like
you
know
the
goggles,
if
you
or
from
the
simpsons,
but
then
on.
You
know
then
on
lumos,
so
it
actually
would
do
something
you
would
you
know,
because
everyone
else
already
does
all
three.
F
So
you
basically
like
that
feature
when
it's
there,
just
you
know,
essentially
gets
ignored
basically
everywhere
else,
essentially
which
like
say,
oh,
it's
kind
of
a
it's
kind
of
unusual,
though,
because
then,
like
I
said,
you're
having
this
feature
it
like
does
nothing
but
in
a
lot
of
places,
and
so
that's
why
I
was
just
in
terms
of
if
anyone
had
ideas
on
approaches.
F
F
You
know
probably
as
much
as
possible
to
get
them
upgrade
and
everything
so
they're
not
out
there
like
that,
although
there
are
going
to
be
some
situations
where
you're
going
to
need
to
retain
compatibility,
because
maybe
you
do
have
to
you
know
you
may
have
systems
on
different
versions
if
you're
exchanging
stuff-
and
so
you
know
you
may
have
some
period
of
time
before
you
can
say
yeah,
I'm
good
across
the
board,
and
so
it's
just
you
know
how
to
manage
that
in
a
way.
F
So
that
way,
you
know
people
are,
you
know,
hey.
You
know.
My
encrypted
data
set
is
no
longer
readable.
You
panic,
you
know.
A
Yeah,
well,
that's
good!
It's
an
interesting
question.
I
think
it's
it's
a
little
bit.
It's
not
obvious
what
the
right
way
to
do
this
is
we've
we've
burned
up
our
usual
time.
Why
don't
we
sort
of
leave
this
as
a
item
to
to
be
thinking
about
and
and
raise
this
as
an
issue
for
the
next
meeting
that
sound
reasonable?
At
this
point,.
F
Yeah,
and
also
I
mean
obviously
those
are-
I
am
on
slack.
When
is
the
do
you
know,
which
date
the
next
meeting
is
going
to
be
specifically
in
august.
F
Yep
17,
yeah
and
yeah,
and-
and
I
am
out
then
and
I'm
going
to
be
in
the
I'll,
be
at
the
bottom
of
the
grand
canyon.
So
I'm
not
going
to
have
any
type
of
connectivity.
Oh
that's
a
problem.
Yes,
so
maybe
be
on
a
raft
down
there,
so
sweet,
but
yes,
okay,
yeah!
So
just
if
anyone
has
ideas,
you
know
feel
free
to
hit
me
up
on
there,
like,
I
said
we'll,
probably
talk
some
more
internally
like
I
said
that
was
the
one
idea
that
came
around.