Re: [PATCH 3/7] net/bonding: change mbuf pool and ring allocation

["Sanford, Robert" <rsanford@akamai.com>]

Hello Connor,

Please see responses inline.

On 12/17/21, 10:44 PM, "Min Hu (Connor)" <humin29@huawei.com> wrote:

> > When the number of used tx-descs (0..255) + number of mbufs in the
> > cache (0..47) reaches 257, then allocation fails.
> > 
> > If I understand the LACP tx-burst code correctly, it would be
> > worse if nb_tx_queues > 1, because (assuming multiple tx-cores)
> > any queue/lcore could xmit an LACPDU. Thus, up to nb_tx_queues *
> > 47 mbufs could be cached, and not accessible from tx_machine().
> > 
> > You would not see this problem if the app xmits other (non-LACP)
> > mbufs on a regular basis, to expedite the clean-up of tx-descs
> > including LACPDU mbufs (unless nb_tx_queues tx-core caches
> > could hold all LACPDU mbufs).
> > 
> I think, we could not see this problem only because the mempool can
> offer much more mbufs than cache size on no-LACP circumstance.
>
> > If we make mempool's cache size 0, then allocation will not fail.
> How about enlarge the size of mempool, i.e., up to 4096 ? I think
> it can also avoid this bug.
> > 
> > A mempool cache for LACPDUs does not offer much additional speed:
> > during alloc, the intr thread does not have default mempool caches
> Why? as I know, all the core has its own default mempool caches ?

These are private mbuf pools that we use *only* for LACPDUs, *one*
mbuf per second, at most. (When LACP link peer selects long timeouts,
we get/put one mbuf every 30 seconds.)

There is *NO* benefit for the consumer thread (interrupt thread
executing tx_machine()) to have caches on per-slave LACPDU pools.
The interrupt thread is a control thread, i.e., a non-EAL thread.
Its lcore_id is LCORE_ID_ANY, so it has no "default cache" in any
mempool.

There is little or no benefit for active data-plane threads to have
caches on per-slave LACPDU pools, because on each pool, the producer
thread puts back, at most, one mbuf per second. There is not much
contention with the consumer (interrupt thread).

I contend that caches are not necessary for these private LACPDU
mbuf pools, but just waste RAM and CPU-cache. If we still insist
on creating them *with* caches, then we should add at least
(cache-size x 1.5 x nb-tx-queues) mbufs per pool.

 
> > Q: Why reserve one additional slot in the rx and tx rings?
> > 
> > A: rte_ring_create() requires the ring size N, to be a power of 2,
> > but it can only store N-1 items. Thus, if we want to store X items,
> Hi, Robert, could you describe it for me?
> I cannot understand why it
> "only store N -1 items". I check the source code, It writes:
> "The real usable ring size is *count-1* instead of *count* to
> differentiate a free ring from an empty ring."
> But I still can not get what it wrote.

I believe there is a mistake in the ring comments (in 3 places).
It would be better if they replace "free" with "full":
"... to differentiate a *full* ring from an empty ring."


> > we need to ask for (at least) X+1. Original code fails when the real
> > desired size is a power of 2, because in such a case, align32pow2
> > does not round up.
> > 
> > For example, say we want a ring to hold 4:
> > 
> >      rte_ring_create(... rte_align32pow2(4) ...)
> > 
> > rte_align32pow2(4) returns 4, and we end up with a ring that only
> > stores 3 items.
> > 
> >      rte_ring_create(... rte_align32pow2(4+1) ...)
> > 
> > rte_align32pow2(5) returns 8, and we end up with a ring that
> > stores up to 7 items, more than we need, but acceptable.
> To fix the bug, how about just setting the flags "RING_F_EXACT_SZ"

Yes, this is a good idea. I will look for examples or test code that
use this flag.

 --
Regards,
Robert Sanford