[dpdk-dev] [PATCH v2] ethdev: document RSS default key and types

[Adrien Mazarguil]

On Wed, Nov 07, 2018 at 03:13:07PM +0000, Ophir Munk wrote:
> 
> > -----Original Message-----
> > From: Adrien Mazarguil [mailto:adrien.mazarguil at 6wind.com]
> > Sent: Wednesday, November 07, 2018 4:06 PM
> > To: Ophir Munk <ophirmu at mellanox.com>
> > Cc: Ferruh Yigit <ferruh.yigit at intel.com>; Andrew Rybchenko
> > <arybchenko at solarflare.com>; dev at dpdk.org; Thomas Monjalon
> > <thomas at monjalon.net>; Asaf Penso <asafp at mellanox.com>; Shahaf
> > Shuler <shahafs at mellanox.com>; Olga Shern <olgas at mellanox.com>
> > Subject: Re: [dpdk-dev] [PATCH v2] ethdev: document RSS default key and
> > types
> > 
> > On Wed, Nov 07, 2018 at 12:39:24PM +0000, Ophir Munk wrote:
> > > > -----Original Message-----
> > > > From: Adrien Mazarguil [mailto:adrien.mazarguil at 6wind.com]
> > > > Sent: Wednesday, November 07, 2018 11:31 AM
> > > > To: Ophir Munk <ophirmu at mellanox.com>
> > > > Cc: Ferruh Yigit <ferruh.yigit at intel.com>; Andrew Rybchenko
> > > > <arybchenko at solarflare.com>; dev at dpdk.org; Thomas Monjalon
> > > > <thomas at monjalon.net>; Asaf Penso <asafp at mellanox.com>; Shahaf
> > > > Shuler <shahafs at mellanox.com>; Olga Shern <olgas at mellanox.com>
> > > > Subject: Re: [dpdk-dev] [PATCH v2] ethdev: document RSS default key
> > > > and types
> > > >
> > > > On Wed, Nov 07, 2018 at 09:23:42AM +0000, Ophir Munk wrote:
> > > > > struct rte_flow_action_rss include fields 'key' and 'types'.
> > > > > Field 'key' is a pointer to bytes array (uint8_t *) which contains
> > > > > the specific RSS hash key.
> > > > > If an application is only interested in default RSS operation it
> > > > > should not care about the specific hash key. The application can
> > > > > set the hash key to NULL such that any PMD uses its default RSS key.
> > > > >
> > > > > Field 'types' is a uint64_t bits flag used to specify a specific
> > > > > RSS hash type such as ETH_RSS_IP (see ETH_RSS_*).
> > > > > If an application does not care about the specific RSS type it can
> > > > > set this field to 0 such that any PMD uses its default type.
> > > > >
> > > > > Signed-off-by: Ophir Munk <ophirmu at mellanox.com>
> > > > > ---
> > > > >  lib/librte_ethdev/rte_flow.h | 9 +++++++--
> > > > >  1 file changed, 7 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/lib/librte_ethdev/rte_flow.h
> > > > > b/lib/librte_ethdev/rte_flow.h index c0fe879..ca9e135 100644
> > > > > --- a/lib/librte_ethdev/rte_flow.h
> > > > > +++ b/lib/librte_ethdev/rte_flow.h
> > > > > @@ -1782,10 +1782,15 @@ struct rte_flow_action_rss {
> > > > >  	 * through.
> > > > >  	 */
> > > > >  	uint32_t level;
> > > > > -	uint64_t types; /**< Specific RSS hash types (see ETH_RSS_*). */
> > > > > +	/**
> > > > > +	 * Specific RSS hash types (see ETH_RSS_*),
> > > > > +	 * or 0 for PMD specific default.
> > > > > +	 */
> > > > > +	uint64_t types;
> > > > >  	uint32_t key_len; /**< Hash key length in bytes. */
> > > > >  	uint32_t queue_num; /**< Number of entries in @p queue. */
> > > > > -	const uint8_t *key; /**< Hash key. */
> > > > > +	/** Hash key, or NULL for PMD specific default key. */
> > > > > +	const uint8_t *key;
> > > >
> > > > I'd suggest to document that on key_len instead. If key_len is
> > > > nonzero, key cannot be NULL anyway.
> > >
> > > The decision if a key/len combination is valid is done in the PMD action
> > validation API.
> > > For example, in MLX5 key==NULL and key_len==40 is accepted.
> > > The combination key==NULL and key_len==0 should always succeeds,
> > however the "must" parameter for RSS default is key==NULL and not
> > key_len==0.
> > 
> > I understand this is how the mlx5 PMD implemented it, but my point is that it
> > makes more sense API-wise to define key_len == 0 as the trigger for a default
> > RSS hash key than key == NULL.
> > 
> > My suggestion is to follow the same trend as memcpy(), mmap(), snprintf()
> > and other well-known functions that take a size when dealing with
> > NULL/undefined pointers. Only size matters! :)
> > 
> 
> Please let's stay backward compatible and consistent with previous dpdk releases where
> key==NULL is used in struct rte_eth_rss_conf (see code snippet in [1]).

And I thought I wouldn't hear again from that structure after ac8d22de2394
("ethdev: flatten RSS configuration in flow API") got rid of it in rte_flow
:)

There is no need to be backward compatible with it particularly if doing so
improves consistency (assuming you agree it does).

Take "rss_hf" versus "types" fields for instance [2]. Unsetting the former
disables RSS completely while unsetting the latter provides default RSS
settings for that flow rule, simply because a RSS action that doesn't do RSS
makes no sense (one could provide a single queue for that).

Regarding "key_len", have a look at this other message [3] in the same
thread which clearly states that i40e expects a zero key length to trigger
default RSS, it doesn't rely on a NULL pointer.

Generally speaking, I'm pushing for zero values in action objects to result
in a safe default behavior. A nonzero key_len with a NULL key may result in
a crash, while the reverse is inherently safer since one doesn't even need
to check the size or pointer validity, e.g.:

 memcpy(target_device_rss_stuff, rss_conf->key, rss_conf->key_len);

What's not to like?

[2] "Re: [PATCH v2 07/15] ethdev: flatten RSS configuration in flow API"
    https://mails.dpdk.org/archives/dev/2018-April/096235.html

[3] "Re: [dpdk-dev] [PATCH v6 07/16] ethdev: flatten RSS configuration in
    flow API"
    https://mails.dpdk.org/archives/dev/2018-May/100128.html

> We should avoid confusing applications with setting key==NULL with legacy RSS and key_len==0 with rte_flows.
> 
> With regard to trends APIs - I was thinking about free() where a NULL pointer is acceptable :)

Yeah, though free() doesn't take a size. On the other hand the behavior of
realloc() is much more interesting when faced with a zero size :) That's
probably one of the few exceptions so I guess my point still stands.

> 
> [1]
> File lib/librte_ethdev/rte_ethdev.h:
> 
> /**
>  * A structure used to configure the Receive Side Scaling (RSS) feature
>  * of an Ethernet port.
>  * If not NULL, the *rss_key* pointer of the *rss_conf* structure points
>  * to an array holding the RSS key to use for hashing specific header
>  * fields of received packets. The length of this array should be indicated
>  * by *rss_key_len* below. Otherwise, a default random hash key is used by
>  * the device driver.
>  *
>  * The *rss_key_len* field of the *rss_conf* structure indicates the length
>  * in bytes of the array pointed by *rss_key*. To be compatible, this length
>  * will be checked in i40e only. Others assume 40 bytes to be used as before.
>  *
>  * ..........
>  */
> struct rte_eth_rss_conf {
> 	uint8_t *rss_key;    /**< If not NULL, 40-byte hash key. */
> 	uint8_t rss_key_len; /**< hash key length in bytes. */
> 	uint64_t rss_hf;     /**< Hash functions to apply - see below. */
> };

-- 
Adrien Mazarguil
6WIND