►
From YouTube: SONiC DASH Workgroup Community Meeting Aug 30 2023
Description
Merge PR422
Discuss PR429
Kumaresh to review PR425 for Prince
Discuss:
PR415 SAI wrong code generated in libsai
PR428 Clarification on Service Tunnel Inbound Processing
P4 pipeline & files Q&A with Marian (resource types, table properties, annotations, bmv2 tables/virtual tables, Rules-Groups-Stages-ENIs, etc...
A
A
But
what
I
wanted
to
point
out
in
case
you
didn't
have
a
chance
to
read
them
is
that
you
know
riff
from
the
Microsoft
Sonic
Team
posted
PR
document
and
some
slides
talking
about
h,
a
work
and
proposal
here,
and
this
happens
in
the
Smart
Switch
Linux
Foundation
call
presented
by
Gohan
and
those
are
on
Thursdays.
One
is
in
the
morning
and
one
is
in
the
evenings.
A
They
alternate
every
other
week,
so
I
pulled
up
the
slides,
I
thought
we
could
take
a
quick
look
at
them,
but
that
happened
and
then
we
I
followed
up
on
PR
360
and
then,
if
everyone's
receiving
the
notes,
it
was
just
what
we
talked
about
last
week
here
so
and
then
the
recordings
are
here
where
you
can
find
them.
So
that's
what
we
did
there
and
before
we
get
started
today.
A
Did
anyone
have
on
the
call
have
a
PR
or
something
they
wanted
to
cover
before
what
I
had
planned
to
do
was
show
you
guys,
basically,
a
quick
overview
of
where
riff
had
posted
his
slides?
If
you
hadn't
seen
him
so,
did
anyone
have
something
first
or
I
know
sod
had
asked
for
a
little
bit
of
time,
but
he
was
waiting
for
someone
familiar
with
the
pipeliner
or
Marion,
but
we
could
always
check
so
did
anyone
else
have
something?
First,.
C
C
C
A
D
B
D
A
E
E
A
B
D
Yeah
so
I
an
API
to
query
how
many
entries
you
can
still
create
in
a
system,
so
the
API
is
per
object
and
also
you
can
have
variations
based
on
attributes.
So
some
some
attributes,
May
May,
Define,
a
more
specific
carving
of
resources,
or
maybe
they
may
say
that
the
size
of
an
entry
is
different.
So
some
of
those
attributes
will
will
be
marked
as
a
resource
type,
which
doesn't
happen
effect
on
the
behavior.
D
B
E
D
B
D
D
B
D
And
there
is
a
list
of
object,
types
that
this
attribute
is
allowed
to
to
be
set
to
like
that
you
you
can
pass
with
this
attribute,
so
it
is
generated
or
it
is
relevant
to
attributes
of
of
PSI
object
types
that
that
are
of
a
that
attribute
is
of
type
object.
Psi
object,
ID,
but
if
you
say
PSI
object,
ID,
it's
not
like
specific
enough
for
validation
checks
right,
because
if
your
attribute
accepts
an
object,
ID,
it
could
be
object.
D
D
E
C
B
E
D
B
E
D
E
B
D
F
B
D
Yeah
yeah,
there
are
inbound
stages.
There
are
outbound
stages.
You
can
pick,
which
type
group
exactly
you
want
to
put
the
rule
in,
because
otherwise
you
would
have
kind
of
not
that
easy
to
use
API
where
you
would
need
to
baste
like
based
on
the
on
the
stage.
You
would
need
to
use
a
different
API
right,
a
different
function
call
and
all
that,
and
all
this
logic
would
need
to
go
into
the
application.
D
B
Okay,
I
think
this
one
I
I
mean
my
I
think
I
have
I,
might
have
incorrect
understanding
about
this
one.
So
a
my
understanding
was
also
looking
at
the
code
that
the
control
plane
creates
an
echo
group
with
rules
and
all
that
and
afterwards
in
a
separate
move.
It
attaches
that
group
to
eni
whether
it
is
inbound
or
which
stage
right.
D
B
B
Example,
while
while
you
are
creating
a
group
right
at
that
moment,
you
don't
care
which,
whether
it
is
inbound
or
outbound
or
which
state.
Now,
once
you
have
treated
a
group
now
you
have,
let's
say
group
ID
now
in
a
separate
API,
for
example,
you
attach
that
group
to
inbound
or
outbound
or,
let's
say
stage,
four
echel
of
the
eni.
D
No
that's
right,
but
look
at
it
from
it's
from
the
B4
or
from
the
pipeline
perspective
right.
You
have
distinct
tables.
You
have
really
like
in
your
pipeline.
You
have
Separate
Tables
that
the
package
should
go
through
and
if
you
just
do
nothing
and
generate
code
for
it,
as
is
you
would
get
an
API
for
each
table.
B
D
D
D
Page
applied
so
like
that,
actually
we
we
probably
need
to
extend
it
to
file
because
size
actually
report.
The
number
of
stages
require
this
five,
so
we
need
to
add
a
four
and
five,
but
really
in
the
behavioral
model.
You
have
only
a
fixed
set
of
tables.
Now
what
you
need
to
kind
of
model
with
that
there
is
your
virtual
teammates,
which
is
this
one
attribute
allows
you
to
do
so.
What
happens
is
that,
like
you
have
10
emis
right?
D
They
belong
to
different
customers
or
to
different
VMS
to
different
v-nets.
They
may
have
completely
different
set
of
those
in
and
out
ACL
groups
attached
to
them,
so
each
eni
may
have
a
unique
set
of
tables,
but
in
the
bmv2
you
have
only
six
tables
and
no
more
so.
What
you
need
to
do
is
when
you
hit
an
Emi
for
a
classification.
So
let's
go
to
the
eni
dash,
that's
probably
Dash
pipeline,
and
there
is
an
eni
table.
D
Here
so
set
Emi
attributes
is
called
when
you
do
an
E
and
I
look.
So
when
you
determine
which
Emi
you
your
packet
belongs
to,
then
you'll
set
your
inbound
outbound
V6
V4
accol
groups,
which
would
be
those
identifiers
of
your
virtual
tables
that
that
will
be
looked
up
when
you,
when
you
do
the
lookup
on
the
on
the
physical
table,
so
you
have
a
probably
I
will
need
to
draw
a
diagram
if
it's
too
complicated.
D
So
when
you
create
a
rule,
you
say
that
rule
belongs
to
a
group.
When
you
create
a
group,
you
attach
it
to,
or
you
say
to
at
which
stage
for
a
particular
eni.
This
group
may
be
applied
and
as
a
result
you
get
like.
So,
first
of
all,
what
we
were
trying
to
do
is,
since
you
have
like
ACL
stages
here
and
ACL
stages
here,
so
it
will
be
6,
P4
tables,
but
really
even
more
because
we
may
add
more
stages.
D
So
we
didn't
want
to
have
that
duplicated
in
the
apis,
because
you
would
get
site
apis
that
look
the
same
just
different
name
of
the
enum
and
different
name
of
the
of
the
function.
So
we
collapsed
those
into
one
API
and
then
you
have
all
these
specifiers
like
the
group
and
then
you
can
attach
the
group
to
ni
and
so
on.
D
Of
those
that
that
is
correct,
yeah,
that
is
correct.
So
the
intention,
by
from
the
user
point
of
view,
is
to
say
which,
which
ACLs
to
look
up
for
this
Emi
and
what
will
really
happen
is
that
we
will,
in
the
behavioral
model
we
will
just
set
in
the
metadata,
those
IDs
and
each
time
we
do
a
lookup
we
will
know
for
which
particular
group
ID.
We
need
to
do
the
look
for.
So
all
the
other
rules
won't
be
relevant.
D
D
B
B
D
Not
find
yeah
right,
it's
not
here,
because
you
can
attach
group
to
any
state
right,
so
rule
doesn't
know
on
which
stage
it
is
executed,
because
you
attach
a
group
to
a
stage
and
you
can
attach
theoretically
I,
don't
know
if
that
really
happens
or
not,
but
it's
not
forbidden.
You
can
attach
group
to
stage
one
in
one
eni
or
to
stage
two
in
a
different
unit.
D
Yeah,
because,
because
for
a
group,
it's
Dynamic,
that's
why
I'm
saying
you
attach
it
you,
you
have
a
fixed
set
of
places
you
can
attach
it
to
which
is
defined
in
the
eni
API.
So
there
is
a
fixed
number
of
places
that
the
group
can
be
attached
to,
but
at
the
create
point
of
creation
of
the
group
or
the
attaching
itself
is
dynamic.
So
in
the
runtime
you
can
attach
the
same
group
to
different
places.
B
Of
yeah
and
then
you
will
go
to
the
eni
API
and
you
will
set
that
group
ID
in
the
eni
attributes
table
right.
So
I'm,
sorry,
that's
that's
what
I
that's
the
understanding
that
I
had,
but
then
why
do?
Why
are
there
Separate
Tables
in
the
pipeline
shouldn't
be
only
a
single
table
in
the
pipeline.
Let's
say:
group
ID,
oh.
B
D
D
B
D
D
B
E
E
B
Says
that
okay,
add
these
rules
in
the
group
and
he
calls
this
API,
okay
and
and
now
in
the
behavioral
model,
I'm
talking
about
the
lips
eye,
for
example,
that
you
generated
there
are
six
of
those
tables.
Okay,
how
will
it
know?
How
will
your
lips,
High
adapter
know
which
table
to
program
the
entry
in
because
you.
D
B
D
Now
you
you
kind
of
can
with
P4
code,
but
then
it
will
be
replicated,
it
will
be
replicated
by
the
compiler,
and
then
there
are
problems
with
that,
because
the
names
generated
are
like
I,
don't
really
control
how
it
will
name
the
tables
so
I
just
created
them
explicitly
myself.
So
there
are
two
ways:
either
you
create
six
tables
or
you
apply
one
table
six
times
and
compiler
will
create
six
tables
for
you.
D
D
D
A
D
A
A
In
case
anyone
wanted
to
go
ahead
and
take
a
look,
the
pr,
the
document,
the
slides
but
he's
really
open
to
feedback,
and
the
audience
here
is
different
than
the
audience
and
the
other
meetings
that
I've
seen
so
I
thought
I
would
bring
it
up
here
and
show
it
to
you
guys
quickly.
So
have
you
guys
been
able
to
see
these
links
or
you
have
these
in
the
notes.
A
A
The
audience
is
different
is
the
thing
not
everybody
here
goes
there
and
vice
versa.
So
I
thought
I
would
just
introduce
it
here
and
let
people
know
they're
here
and
then
there
is
a
if
they
want
to
go
through
the
Deep
review.
I
guess
you
could
look
at
the
Pierre,
the
docs,
the
slides
and
then
go
to
that
call.
I!
Guess
if
you
want
to
go
into
the
in-depth
review,
I
guess:
that's
all
I
was
trying
to
say
if
anyone's
interested.
A
E
A
C
C
D
D
C
E
A
C
A
C
F
Basically,
what's
happening
right
now,
this
is
just
a
compiler
warning,
but
under
the
hood,
the
oil,
the
ID,
is
to
just
lower
16
16
piece
from
64
bits,
yeah,
so
I'm
not
sure.
You
know
whether
this
have
some
like
impact
internally,
whether
in
the
BMV
tools,
the
streets
or
whether
in
the
side
drift
so
I
actually
need
someone
with
like
deeper
knowledge.
F
D
D
Which
is
less
than
64,
but
it's
16
bits
or
something
like
that
and
you
you
already
know
what
the
type
is.
You
don't
need
the
whole
object.
Id.
You
only
need
to
pass
the
year
ID
irrelevant
to
bnv2,
but
I
I
will
check.
Maybe
there
is
a
way
to
do
it
without
These
Warnings,
just
explicitly
split
the
the
object.
Id.
D
F
F
F
E
A
A
A
E
E
The
outbound
processing
creates
a
entry
in
the
flow
table
for
matching
on
subsequent
packets
in
that
outbound
flow,
and
it
also
creates
an
inbound
flow
entry
for
packets
coming
from
the
other
direction,
so
here
I
believe
the
the
packets
coming
from
the
other
direction
or
IPv6
and
the
packets
that
are
going
out,
outbound,
originally
rpv4,
so
I'm
trying
to
understand.
If
this
description
is
saying
or
implying
that
yeah
we
create
a
V4
entry
for
outbound
and
a
V6
entry
for
inbound.
E
E
Then
I
mean
it's
just
different
processing.
Usually
you
you
pop
off
the
outer
header,
you're
left
with
the
remaining
header,
and
you
do
a
flow
lookup
on
it.
This
would
be
you'd
pop
off.
An
outer
header,
you'd
know
somehow
to
skip
the
middle
header
and
then
use
the
inner
header
for
the
flow
lookup
and
like
how
do
you
know
to
skip
the
middle
like
mm-hmm,
yeah,.
E
So
I
think
in
the
hld
in
this
section,
there's
lots
of
good
examples
like
with
specific,
like
packets
and
configs
and
IP
addresses,
and
a
good
description
of
a
step-by-step
processing
for
the
outbound
and
and
what
happens
at
how
you
arrive
at
the
final
outcome
packet.
But
the
inbound
description
is
is
much
more
brief
and
it
would
be
useful
I
think
if
there
were
inbound
descriptions
as
rich
as
the
outbound.