Hello Connor,
Thank you for the questions and comments. I will repeat the questions, followed
by my answers.
Q: Could you be more detailed, why is mbuf pool caching not needed?
A: The short answer: under certain conditions, we can run out of
buffers from that small, LACPDU-mempool. We actually saw this occur
in production, on mostly-idle links.
For a long explanation, let's assume the following:
1. 1 tx-queue per bond and underlying ethdev ports.
2. 256 tx-descriptors (per ethdev port).
3. 257 mbufs in each port's LACPDU-pool, as computed by
bond_mode_8023ad_activate_slave(), and cache-size 32.
4. The "app" xmits zero packets to this bond for a long time.
5. In EAL intr thread context, LACP tx_machine() allocates 1 mbuf
(LACPDU) per second from the pool, and puts it into LACP tx-ring.
6. Every second, another thread, let's call it the tx-core, calls
tx-burst (with zero packets to xmit), finds 1 mbuf on LACP tx-ring,
and underlying ethdev PMD puts mbuf data into a tx-desc.
7. PMD tx-burst configured not to clean up used tx-descs until
there are almost none free, e.g., less than pool's cache-size *
CACHE_FLUSH_THRESH_MULTIPLIER (1.5).
8. When cleaning up tx-descs, we may leave up to 47 mbufs in the
tx-core's LACPDU-pool cache (not accessible from intr thread).
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).
If we make mempool's cache size 0, then allocation will not fail.
A mempool cache for LACPDUs does not offer much additional speed:
during alloc, the intr thread does not have default mempool caches
(AFAIK); and the average time between frees is either 1 second (LACP
short timeouts) or 10 seconds (long timeouts), i.e., infrequent.
--------
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,
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.
--------
Q: I found the comment for BOND_MODE_8023AX_SLAVE_RX_PKTS is
wrong, could you fix it in this patch?
A: Yes, I will fix it in the next version of the patch.
--
Regards,
Robert Sanford
On 12/16/21, 4:01 AM, "Min Hu (Connor)" <humi...@huawei.com> wrote:
Hi, Robert,
在 2021/12/16 2:19, Robert Sanford 写道:
> - Turn off mbuf pool caching to avoid mbufs lingering in pool caches.
> At most, we transmit one LACPDU per second, per port.
Could you be more detailed, why does mbuf pool caching is not needed?
> - Fix calculation of ring sizes, taking into account that a ring of
> size N holds up to N-1 items.
Same to that, why should resvere another items ?
>
By the way, I found the comment for BOND_MODE_8023AX_SLAVE_RX_PKTS is
is wrong, could you fix it in this patch?
> Signed-off-by: Robert Sanford <rsanf...@akamai.com>
> ---
> drivers/net/bonding/rte_eth_bond_8023ad.c | 14 ++++++++------
> 1 file changed, 8 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/net/bonding/rte_eth_bond_8023ad.c
b/drivers/net/bonding/rte_eth_bond_8023ad.c
> index 43231bc..83d3938 100644
> --- a/drivers/net/bonding/rte_eth_bond_8023ad.c
> +++ b/drivers/net/bonding/rte_eth_bond_8023ad.c
> @@ -1101,9 +1101,7 @@ bond_mode_8023ad_activate_slave(struct rte_eth_dev
*bond_dev,
> }
>
> snprintf(mem_name, RTE_DIM(mem_name), "slave_port%u_pool",
slave_id);
> - port->mbuf_pool = rte_pktmbuf_pool_create(mem_name, total_tx_desc,
> - RTE_MEMPOOL_CACHE_MAX_SIZE >= 32 ?
> - 32 : RTE_MEMPOOL_CACHE_MAX_SIZE,
> + port->mbuf_pool = rte_pktmbuf_pool_create(mem_name, total_tx_desc, 0,
> 0, element_size, socket_id);
>
> /* Any memory allocation failure in initialization is critical
because
> @@ -1113,19 +1111,23 @@ bond_mode_8023ad_activate_slave(struct
rte_eth_dev *bond_dev,
> slave_id, mem_name, rte_strerror(rte_errno));
> }
>
> + /* Add one extra because ring reserves one. */
> snprintf(mem_name, RTE_DIM(mem_name), "slave_%u_rx", slave_id);
> port->rx_ring = rte_ring_create(mem_name,
> - rte_align32pow2(BOND_MODE_8023AX_SLAVE_RX_PKTS),
socket_id, 0);
> + rte_align32pow2(BOND_MODE_8023AX_SLAVE_RX_PKTS + 1),
> + socket_id, 0);
>
> if (port->rx_ring == NULL) {
> rte_panic("Slave %u: Failed to create rx ring '%s':
%s\n", slave_id,
> mem_name, rte_strerror(rte_errno));
> }
>
> - /* TX ring is at least one pkt longer to make room for marker packet. */
> + /* TX ring is at least one pkt longer to make room for marker packet.
> + * Add one extra because ring reserves one. */
> snprintf(mem_name, RTE_DIM(mem_name), "slave_%u_tx", slave_id);
> port->tx_ring = rte_ring_create(mem_name,
> - rte_align32pow2(BOND_MODE_8023AX_SLAVE_TX_PKTS + 1),
socket_id, 0);
> + rte_align32pow2(BOND_MODE_8023AX_SLAVE_TX_PKTS + 2),
> + socket_id, 0);
>
> if (port->tx_ring == NULL) {
> rte_panic("Slave %u: Failed to create tx ring '%s':
%s\n", slave_id,
>
