Re: [PATCH v8 30/30] packet: Add support for multi-vector packets

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 5137 bytes --]

On Thu, Aug 14, 2025 at 11:02:57AM +0200, Laurent Vivier wrote:
> On 07/08/2025 08:17, David Gibson wrote:
> > On Tue, Aug 05, 2025 at 05:46:28PM +0200, Laurent Vivier wrote:
> > > The packet pool was previously limited to handling packets contained
> > > within a single buffer.
> > > 
> > > This patch extends the packet pool to support iovec array,
> > > allowing a single logical packet to be composed of multiple iovec.
> > > 
> > > To accommodate this, the storage format within the pool is modified.
> > > For a multi-vector packet, a header entry is now stored first with
> > > iov_base = NULL and iov_len holding the number of subsequent
> > > vectors. The actual data vectors are then stored in the following
> > > pool slots.
> > > 
> > > The packet_add_do() and packet_get_do() functions are updated to
> > > manage this new format for storing and retrieving packets. The
> > > pool_full() check is also adjusted to ensure there is enough
> > > space for all vectors of a new packet before adding it.
> > > 
> > > Signed-off-by: Laurent Vivier <lvivier@redhat.com>
> > > ---
> > >   packet.c | 50 +++++++++++++++++++++++++++++++++-----------------
> > >   packet.h |  2 +-
> > >   tap.c    |  4 ++--
> > >   3 files changed, 36 insertions(+), 20 deletions(-)
> > > 
> > > diff --git a/packet.c b/packet.c
> > > index 4b93688509a4..d697232d951a 100644
> > > --- a/packet.c
> > > +++ b/packet.c
> > > @@ -90,12 +90,13 @@ static int packet_check_range(const struct pool *p, const char *ptr, size_t len,
> > >   /**
> > >    * pool_full() - Is a packet pool full?
> > >    * @p:		Pointer to packet pool
> > > + * @data:	check data can fit in the pool
> > >    *
> > > - * Return: true if the pool is full, false if more packets can be added
> > > + * Return: true if the pool is full, false if data can be added
> > >    */
> > > -bool pool_full(const struct pool *p)
> > > +bool pool_full(const struct pool *p, const struct iov_tail *data)
> > 
> > Given the slightly changed semantics, I wonder if 'pool_can_fit()'
> > might be a better name now.
> 
> okay
> 
> > 
> > >   {
> > > -	return p->count >= p->size;
> > > +	return p->count + data->cnt  + (data->cnt > 1) >= p->size;
> > 
> > This test is only correct if data is already pruned.  As I've said
> > elsewhere, it might be worth changing to the assumption that iov_tails
> > are pruned everywhere outside the iov_tail internal handling.
> > 
> > Oh.. also I think the new check is off by one (in the relatively safe
> > direction).  It will say there's no room when there is just exactly
> > enough room.
> 
> Could you explain why you think it's off by 1?

Loosely, because the old version needed to check if the pool was
_already_ full, but the new version needs to check if it _will_ be
full with the addition.

Suppose there's exactly one free slot in the pool, say p->count==99,
p->size==100.  We attempt to add a single buffer packet, data->cnt==1.

	99 + 1 + 0 >= 100

But it should (just) fit.

> > >   }
> > >   /**
> > > @@ -108,11 +109,9 @@ bool pool_full(const struct pool *p)
> > >   void packet_add_do(struct pool *p, struct iov_tail *data,
> > >   		   const char *func, int line)
> > >   {
> > > -	size_t idx = p->count;
> > > -	const char *start;
> > > -	size_t len;
> > > +	size_t idx = p->count, i, offset;
> > > -	if (pool_full(p)) {
> > > +	if (pool_full(p, data)) {
> > >   		debug("add packet index %zu to pool with size %zu, %s:%i",
> > >   		      idx, p->size, func, line);
> > >   		return;
> > > @@ -121,18 +120,30 @@ void packet_add_do(struct pool *p, struct iov_tail *data,
> > >   	if (!iov_tail_prune(data))
> > >   		return;
> > > -	ASSERT(data->cnt == 1); /* we don't support iovec */
> > > +	if (data->cnt > 1) {
> > > +		p->pkt[idx].iov_base = NULL;
> > > +		p->pkt[idx].iov_len = data->cnt;
> > > +		idx++;
> > > +	}
> > > -	len = data->iov[0].iov_len - data->off;
> > > -	start = (char *)data->iov[0].iov_base + data->off;
> > > +	offset = data->off;
> > > +	for (i = 0; i < data->cnt; i++) {
> > > +		const char *start;
> > > +		size_t len;
> > > -	if (packet_check_range(p, start, len, func, line))
> > > -		return;
> > > +		len = data->iov[i].iov_len - offset;
> > > +		start = (char *)data->iov[i].iov_base + offset;
> > > +		offset = 0;
> > > -	p->pkt[idx].iov_base = (void *)start;
> > > -	p->pkt[idx].iov_len = len;
> > > +		if (packet_check_range(p, start, len, func, line))
> > > +			return;
> > > -	p->count++;
> > > +		p->pkt[idx].iov_base = (void *)start;
> > > +		p->pkt[idx].iov_len = len;
> > > +		idx++;
> > 
> > Hm. Isn't the above equivalent to iov_tail_clone()?  Is calling
> > packet_check_range() on each chunk the only reason for open-coding it
> > here?
> 
> Yes, I think the code is clearer like this. And it avoids to scan the iovec
> array twice (with the offset management).

Ok.

-- 
David Gibson (he or they)	| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you, not the other way
				| around.
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]