Re: [PATCH net-next v6 06/15] ethtool: netlink bitset handling

[Michal Kubecek]

On Thu, Jul 04, 2019 at 10:04:35AM +0200, Jiri Pirko wrote:
> Wed, Jul 03, 2019 at 08:18:51PM CEST, mkubecek@xxxxxxx wrote:
> >On Wed, Jul 03, 2019 at 01:49:33PM +0200, Jiri Pirko wrote:
> >> Tue, Jul 02, 2019 at 01:50:09PM CEST, mkubecek@xxxxxxx wrote:
> >> >+Compact form: nested (bitset) atrribute contents:
> >> >+
> >> >+    ETHTOOL_A_BITSET_LIST	(flag)		no mask, only a list
> >> >+    ETHTOOL_A_BITSET_SIZE	(u32)		number of significant bits
> >> >+    ETHTOOL_A_BITSET_VALUE	(binary)	bitmap of bit values
> >> >+    ETHTOOL_A_BITSET_MASK	(binary)	bitmap of valid bits
> >> >+
> >> >+Value and mask must have length at least ETHTOOL_A_BITSET_SIZE bits rounded up
> >> >+to a multiple of 32 bits. They consist of 32-bit words in host byte order,
> >> 
> >> Looks like the blocks are similar to NLA_BITFIELD32. Why don't you user
> >> nested array of NLA_BITFIELD32 instead?
> >
> >That would mean a layout like
> >
> >  4 bytes of attr header
> >  4 bytes of value
> >  4 bytes of mask
> >  4 bytes of attr header
> >  4 bytes of value
> >  4 bytes of mask
> >  ...
> >
> >i.e. interleaved headers, words of value and words of mask. Having value
> >and mask contiguous looks cleaner to me. Also, I can quickly check the
> >sizes without iterating through a (potentially long) array.
> 
> Yeah, if you are not happy with this, I suggest to introduce
> NLA_BITFIELD with arbitrary size. That would be probably cleanest.

There is still the question if it it should be implemented as a nested
attribute which could look like the current compact form without the
"list" flag (if there is no mask, it's a list). Or an unstructured data
block consisting of u32 bit length and one or two bitmaps of
corresponding length. I would prefer the nested attribute, netlink was
designed to represent structured data, passing structures as binary goes
against the design (just looked at VFINFO in rtnetlink few days ago,
it's awful, IMHO).

Either way, I would still prefer to have bitmaps represented as an array
of 32-bit blocks in host byte order. This would be easy to handle in
kernel both in places where we have u32 based bitmaps and unsigned long
based ones. Other options seem less appealing:

  - u8 based: only complicates processing
  - u64 based: have to care about alignment
  - unsigned long based: alignment and also problems with 64-bit kernel
    vs. 32-bit userspace

> >> This is quite complex and confusing. Having the same API for 2 APIs is
> >> odd. The API should be crystal clear, easy to use.
> >> 
> >> Why can't you have 2 commands, one working with bit arrays only, one
> >> working with strings? Something like:
> >> X_GET
> >>    ETHTOOL_A_BITS (nested)
> >>       ETHTOOL_A_BIT_ARRAY (BITFIELD32)
> >> X_NAMES_GET
> >>    ETHTOOL_A_BIT_NAMES (nested)
> >> 	ETHTOOL_A_BIT_INDEX
> >> 	ETHTOOL_A_BIT_NAME
> >> 
> >> For set, you can also have multiple cmds:
> >> X_SET  - to set many at once, by bit index
> >>    ETHTOOL_A_BITS (nested)
> >>       ETHTOOL_A_BIT_ARRAY (BITFIELD32)
> >> X_ONE_SET   - to set one, by bit index
> >>    ETHTOOL_A_BIT_INDEX
> >>    ETHTOOL_A_BIT_VALUE
> >> X_ONE_SET   - to set one, by name
> >>    ETHTOOL_A_BIT_NAME
> >>    ETHTOOL_A_BIT_VALUE
> >
> >This looks as if you assume there is nothing except the bitset in the
> >message but that is not true. Even with your proposed breaking of
> >current groups, you would still have e.g. 4 bitsets in reply to netdev
> >features query, 3 in timestamping info GET request and often bitsets
> >combined with other data (e.g. WoL modes and optional WoL password).
> >If you wanted to further refine the message granularity to the level of
> >single parameters, we might be out of message type ids already.
> 
> You can still have multiple bitsets(bitfields) in single message and
> have separate cmd/cmds to get string-bit mapping. No need to mangle it.

Let's take a look at what it means in practice, the command is

  ethtool --set-prif-flags eth3 legacy-rx on

on an ixgbe card. Currently, ethtool (from the github repository) does

------------------------------------------------------------------------
ETHTOOL_CMD_SETTINGS_SET (K->U, 68 bytes)
    ETHTOOL_A_HEADER
        ETHTOOL_A_DEV_NAME = "eth3"
    ETHTOOL_A_SETTINGS_PRIV_FLAGS
        ETHTOOL_A_BITSET_BITS
            ETHTOOL_A_BITS_BIT
                ETHTOOL_A_BIT_NAME = "legacy-rx"
                ETHTOOL_A_BIT_VALUE

NLMSG_ERR (K->U, 36 bytes) err = 0
------------------------------------------------------------------------

If we had only compact form (or some of the NLA_BITFIELD solutions we
are talking about), you would need

------------------------------------------------------------------------
ETHTOOL_CMD_STRSET_GET (U->K, 52 bytes)
    ETHTOOL_A_HEADER
        ETHTOOL_A_DEV_NAME = "eth3"
    ETHTOOL_A_STRSET_STRINGSETS
        ETHTOOL_A_STRINGSETS_STRINGSET
            ETHTOOL_A_STRINGSET_ID = 2 (ETH_SS_PRIV_FLAGS)

ETHTOOL_CMD_STRSET_GET_REPLY (K->U, 128 bytes)
    ETHTOOL_A_HEADER
        ETHTOOL_A_DEV_INDEX = 9
        ETHTOOL_A_DEV_NAME = "eth3"
    ETHTOOL_A_STRSET_STRINGSETS
        ETHTOOL_A_STRINGSETS_STRINGSET
            ETHTOOL_A_STRINGSET_ID = 2 (ETH_SS_PRIV_FLAGS)
            ETHTOOL_A_STRINGSET_COUNT = 2
            ETHTOOL_A_STRINGSET_STRINGS
                ETHTOOL_A_STRINGS_STRING
                    ETHTOOL_A_STRING_INDEX = 0
                    ETHTOOL_A_STRING_VALUE = "legacy-rx"
                ETHTOOL_A_STRINGS_STRING
                    ETHTOOL_A_STRING_INDEX = 1
                    ETHTOOL_A_STRING_VALUE = "vf-ipsec"

NLMSG_ERR (K->U, 36 bytes) err = 0

ETHTOOL_CMD_SETTINGS_SET (K->U, 64 bytes)
    ETHTOOL_A_HEADER
        ETHTOOL_A_DEV_NAME = "eth3"
    ETHTOOL_A_SETTINGS_PRIV_FLAGS
        ETHTOOL_A_BITSET_SIZE = 2
        ETHTOOL_A_BITSET_VALUE = 00000001
        ETHTOOL_A_BITSET_MASK = 00000001

NLMSG_ERR (K->U, 36 bytes) err = 0
------------------------------------------------------------------------

That's an extra roundtrip, lot more chat and the SETTINGS_SET message is
only 4 bytes shorter in the end. And we can consider ourselves lucky
this NIC has only two private flags. Or that we didn't need to enable or
disable a netdev feature (56 bits) or link mode (69 bits and growing).

We could reduce the overhead by allowing STRSET_GET query to only ask
for specific string(s) but there would still be the extra roundtrip
which I dislike in the ioctl interface. Florian also said in the v5
discussion that he would like if it was possible to get names and data
together in one request.

Michal