►
From YouTube: CDS Jewel -- CephFS starter tasks
B
This
is
the
command
in
the
MDS
that
lets
you
lets.
You
just
tell
it
to
dump
out
the
state
of
the
entire
cache
that
it
already
has
in
memory
and
that
can
either
come
out
over
standard
outer
to
it
to
a
file
on
on
desk
and
right
now,
all
its
great
course.
You
can
help
it
up
to
cash
and
dumped
everything
that
we
have.
B
But
then,
then,
if
you
want
to
do
something
specific
about
that,
you
need
to
go
through
and
pull
out
all
the
information
you're
interested
in.
So
we
have
a
ticket
to
do
to
dump
the
cash
for
that
a
particular
client
is
involved
with
or
that
an
inode,
or
that
is
one
particular
I
know'd
or
that
lives
underneath
some
specific
path,
and
this
isn't
too
complicated.
B
If
you
are
interested
in
doing
it,
there
is
a
single
function
called
dump
cash
in
the
md
cash
comes
class
and
right
now
it
just
runs
through
everything
and
if
you're
interested
in
wanting
to
get
stick
it
to
just
go
into
that
function
and
add
a
couple
of
parameters
to
it,
based
on
or
maybe
you'd
clone
it
or
something,
but
you'd.
Basically,
take
in
the
through
the
admin
socket
the
state
you're
interested
in
and
then
just
go
through
the
dub
cash
function
and
filter
in
the
loops.
Based
on
these
things,
I
guess
another.
A
B
B
A
Yeah,
the
arts
right
yeah,
the
I
mean
all
the
code
exists
for
touching
them
for
basically,
but
some
is
not
wrapped
up
in
am
in
an
actual.
You
know.
Ls
come
on
that
will
probably
eat
a
natural
thing
to
do
when
we
might
be
heading
out
of
the
colors
of
the
am
of
the
metadata
objects,
because
we're
going
to
need
something
that
knows
how
to
walk
the
the
frame
fragments
in
a
directory,
all
I
suppose
if
this
tool
was
limited
to
just
listing
an
individual
turf,
Ragan
didn't
have
to
know
about
factories,
then
it's
yeah.
B
So
if
you're
interesting
this
one,
you
should
come
talk
to
us,
but
basically
what
this
is
gonna
have
to
do
is:
go
eat
an
object
off
disc
and
I
and
I.
Don't
remember
how
many
help
and
functions
we
have
for
all
of
this,
but
to
go,
read
an
object
out
of
regos
and
then
just
parse
it
and
there's
parsing
code
in
various
places.
You'd
probably
have
to
have
to
make
some
bit
a
little
neater.
B
B
You
know
we
want
to
dump
some
or
a
lot
of
times.
It's
saying
that
some
complicated
thing
is
false
and
forth.
At
least
it's
something.
If
it's
false
is,
there
is
all
of
us
a
complicated
set
of
set
of
situations,
so
it
would
be
really
good
to
get
more
detail
about
what's
going
on
in
the
system
when
these
asserts
fail,
and
we
make
some
attempt
to
do
that
by.
B
But
if
we
assert
out,
then
we
do
dump
them,
but
John
Jones
listed
here,
a
dump
of
Demi
cash,
a
dump
of
the
ops
in
flight
and
then
the
and
then
options
to
add
information
to
the
central
monitor
log
or
to
tell
them
MDS
monitor
that
we're
dying
so
that
we
don't
have
to
wait
through
the
timeout.
While
the
monitors
detect
that
were
dead
before
the
failover
happens
or
just
the
system
gets
marked.
Lanky.
B
And
so
to
this,
you'd
have
to
write
from
whole
cloth,
but
on
that
not
from
whole
cloth.
So,
as
I
mentioned,
we
have
a
function
called
md
cat
or
a
function
to
dump
the
mb
cache
cache.
So
you
just
have
to
have
to
call
that
there's
a
dumb
papa
option
flight
command
in
the
app
tracker
that
you
have
to
call.
B
Have
to
send
yours
figure
out
how
to
send
the
messages,
but
we
have
helpers
for
all
that
sort
of
the
part
of
this.
It
would
be
difficult,
I
think,
is
too
it
is
learning
about
the
infrastructure
involved
in
and
how
the
generic
asserts,
work
and
seth
and
then
going
through
and
figuring
out
a
way
to
to
do
these
things,
while
blocking
other
progress
on
which
suddenly
makes
me
think
that
maybe
this
one's
harder
than
I
thought,
but
so.
A
A
Yeah
you
can,
you
can
send
you
you
can't
receive,
because
all
you
you'll
be
in
the
middle
of
our
dispatch
leave,
so
you
can
enter
you
not
hang
someone
seeing
that
message,
but
yeah
it's
be.
The
the
sending
of
the
message
to
the
mom
is
basically
the
same
path,
as
already
happens
when
you
call
damaged
yeah
but
becomes
or
send
and
wait
or
whatever
it's
called,
and
that
pretty
much
Milan
works
well
enough,
while
you're
holding
the
MDS
lock.
B
Okay
and
then
this
one
this
last
one
that
I
had
in
the
MDS
I'm,
actually
not
sure
how
difficult
it
would
be,
but
I
think
would
be
interesting
for
someone
who
wanted
to
get
in
in
deep,
which
is
to
support
catch
eviction,
and
this
is
so
the
md
cash
assets
and
memory
cache.
But
it's
really
difficult
to
tell
if
something
is
been
removed
from
that
cash
or
not,
which
can
be
a
useful
thing
to
check
to
want
to
check
when
you're
running
white
box
tests
of
the
stuff
of
that
system.
B
So
I
don't
think
this
would
work,
or
this
wouldn't
work
if
clients
had
had
to
file
open
or
something.
But
if
clients
who
weren't
actually
using
the
file,
but
just
had
you
know
extra
extra
caps
on
it
because
they
like
to
keep
caps
around
or
if
no
clients
had
it
in
use,
then
we
want
to
be
able
to
just
dump
inodes
out
of
the
cash
then
flush
them
to
this,
give
that's
necessary
or
whatever
and
I'd
like
to
command
via
the
admin
socket
pretty
bumpers.
B
Ology
sweet
would
write
a
test
that
does
specific
things
with
I
nodes
and
then
kicks
them
out
of
cash
and
then
does
something
else
to
make
sure
that
the
state
of
the
inode
or
that
the
inode
is
in
the
right
state
and
that
we
don't
have
some
bug
in
how
we're
doing
right
out
and
and
and
then
reading
off
of
radios,
that's
being
masked
by
it.
Just
staying
in
memory
all
the
time
or
something.
B
So
the
demon,
all
the
demons,
have
something
called
the
I
have
their
emptier
admin
socket.
You
can
issue
sort
of
debugging
and
commands
to
okay,
the
most
common
uses
right
now
or
for
like
dumping,
there's
something
there's
a
knob
tracker
in
the
MDS
in
the
OSD
now
which
every
time
there's
a
message
from
some
other
demon
or
from
a
client
or
whatever.
Then
it
creates
an
OP
and
you
can
track
these
and
so
there's
a
command
dump
Hopsin
flight.
That
shows
you
everything
that's
happening
on
the
system
right
now.
B
B
Let's
see,
help
me
out
John,
which
was
which
they're
actually
in
the
atom
socket,
because
they
all
should
be,
but
I
know
we
just
moved
a
bunch
of
them
in
there
and
I'm,
not
sure
everything
is
so
like
I
think
we
have
commands
for
migrating
subtrees
between
metadata
servers,
which
is
something
you
shouldn't
really
do,
except
when
you're
trying
to
set
up
a
specific
scenario.
Yeah.
A
We
have
there's
my
right
directory
and
also
I
just
readded.
The
split
emerged,
the
directory
fragmentation
right.
Almost
all
a
couple
of
old
legacy,
MDS
tell
ones
that
haven't
been
reinstated
yet,
but
yeah
I
mean
they're
they're
entirely
or
not
singer
and
use
with
mr.
we
want
to
do
actually.
Maybe
we
should
start
practicing
these
things
with
debug
or
something
like
that.
A
B
Or
requiring
it
that
you
start
up
the
demon
in
a
debug
mode
or
something
to
to
make
use
of
them
anyway,
yeah,
so
those
are
sort
of
like
flip
options
for
twiddling
the
internal
state
of
the
MDS.
Without
that
you
wouldn't
want
to
do
this
because
you're,
because
it's
running
in
a
cluster
somewhere,
you
do
it
because
you
have
a
specific
goal
involved
with
trying
to
test
this
visit
feature,
and
this
would
be
one
of
those
okay.
B
B
B
Sort
of
them,
like
you
kind
of
like
the
nun,
we
need
to
go
down
into
our
actual
client
of
secrecy
files
as
well
to
deal
with
this.
If
you
do
it's
not
hard,
like
the
like
I'm
looking
at
the
the
ll
read
function,
which
I
think
is
what
are
actually
don't
remember,
which
vary
that
lips
FSU's
is
about
like
our
ll
read
function
is,
is
blocking
but
sort
of
the
blocking
stuff
is
internally
acing.
So
it's
not
hard,
so
it
wouldn't
be
hard
to
like
expose
that
from
just
one
level
down.
B
So
that
would
get
you
on
the
client
side
and
sort
of
give
you
a
tour
of
the
important
interfaces
involved
there
thanks.
What
is
the
user
space
client
has,
along
with
most
of
the
other
set
processes,
has
something
called
proof:
counters
performance
account
thats
short
for
performance
counters,
which
you
can
access
to
the
admin
socket
and
to
get
their
states
and
they're
just
counters
of
averages
or
sums
or
current
values,
or
whatever
that's
taken
number
for
355.
B
B
The
big
one
that
I
think
that
we
should
have
is
the
number
of
requests
which
are
outstanding
to
the
MDS.
But
anybody
who
is
interested
in
getting
to
know
the
client
side
code,
just
like
browsing
just
browsing
through
that
code
and
I'm
looking
for
important
data
structures,
would
see
all
kinds
of
things
that
that
would
be
used.
That
could
be
so
to
put
in
a
structure
like
that,
so
I
started
going
through
it
last
night.
It
was
like
okay,
so
the
size
of
the
cache
of
the.
I
knows
that
we
have
internally
be
sigh.
B
B
When
you
invoke
it
and
there's
no
indication
of
what
of
what's
going
on
and
why
it's
not
proceeding
to
mount
and
every
self,
and
we
actually
get
someone
who,
on
the
mailing
was
to
you
know,
ran
through
the
South
quick
start
and
didn't
set
up
an
MDS
cluster
and
tries
to
mount
with
SEF
fuse
and
doesn't
know
why.
That's
not
working.
B
B
B
Now
right
now,
the
FGS
is
responsible
for
writing
this
out,
in
which
you
know,
works.
Fine
and
the
MDS
needs
to
be
able
to
write
it
out.
In
case
you
have
an
old
client
connecting,
but
it
does
feed
that
in
a
lot
of
common
use
cases,
some
of
which
people
have
complained
to
us
about
one
at
what
it's
like
something
like
our
sink.
B
Then,
if
your
are
seeing
a
bunch
of
small
files,
then
you
have
to
rights
to
disk
for
every
file,
one
when
the
client
writes
out
the
file
disk
event,
one
when
the
MBS
writes
the
back-trace
to
that
to
that
first
object
and
that's
sort
of
annoying,
and
since
the
client
is
already
tends
to
access
the
first
object.
Whatever
it
creates
a
new
file,
it
would
be
convenient
if
the
if
the
client
was
writing
that
factories
out
for
us,
so
I
think
the
way
to
do
this
without
making
everything
super
complicated
is
just
too
it's
just.
B
B
We
could
try
you
something
more
complicated
with
by
making
it
a
piece
of
view
of
the
full
capability
system
and
I'm,
not
sure.
If
that
would
have
any
real
advantages
for
us
like
in
the
future
on,
we
could
try
and
get
the
client
to
do
to
write
out
extra
back
traces
for
us,
not
just
the
most
recent,
like
not
just
on,
create
and
I.
Think
that
might
be
here
with
you
with
a
full
capability,
but
I'm
not
really
sure.
A
So
you'll
in
time,
if
with
that
mechanism,
generally,
you
only
get
the
client
writing
out
the
back-trace
if
it
make,
or
rather
you
only
save
yourself
from
having
to
ride
the
back-trace
from
the
MDS.
If
the
client
writes
back
its
cast,
indicating
that
it's
written
the
back-trace
before
your
create
falls
off
the
end
of
the
journal
and
has
to
get
has
to
get
written
now
so
I
wonder
if
we
own
so
we
could.
A
B
So
I
don't
know,
and
also
you
just
assume
that
ended
generally
speaking,
the
decree
isn't
going
to
fall
off
that
quickly
compared
to
the
clients
accesses.
That
might
be
a
false
assumption.
In
particular,
do
we
actually
mean
the
Creator
fall
off
or
I?
Think
we
just
need
to
keep
the
I
toted
it's
not
when
it
create
false
off
of
what.
As
long
as
the
I
knows,
you
can
cash
right.
B
Okay-
and
we
have
two
minutes-
left,
okay,
so
I
guess
we
run
through
the
treaty
quickly
number
of
3
to
4
4.
Is
that
so
we
have
something
on
FS,
al
south
of
abstraction
layer
for
the
NFS
ganesha
demon
project,
which
speaks
which
binds
until
it's
F
of
s
and
but
but
we
don't
test
them
at
all.
Unlike
the
/
Samba
blind
date,
or
anything
else
so
and
I
got
a
report
recently
that
they
were
a
little
bit
broken
so
be
great
if
we
could
test
them
into
tholla
G.
B
So
this
one
would
be
more
about
doing
the
work
to
set
up
to
ecology
tests
that
would
like
install
Ganesha
and
also
possible,
and
also
to
do
that.
Then
we
need
to
like
you
might
need
to
get
help
from
us
and
lab
access
to
like
set
up
a
good
Isha,
get
builder
all
I'm,
not
sure
about
that.
But
it
would
be
great
if
someone
good
like
right
to
write
some
do
value
tasks
that
install
and
test
Kamisha
and
you
could
like
for
the
test.
B
It
would
like
you
would
do
a
lot
of
stuff
with
the
southwest
code,
but
it
would
be
mostly
about,
like
the
compile
time,
trying
to
split
that
up
into
a
an
internal
MBS
types
that
are
only
used
by
the
MDS
and
then
all
you
need
to
link
with
it
and
in
two
types
that
are
shared
with
a
client,
and
that
would
be
good
because
it
would
make
build
to
a
faster
at
a
lot
of
cases
and
it
was
baked
linking
and
packaging
and
various
other
things
that
we
run
into
across
the
years.
A
lot
simpler.