From: Ryan Wolfe <rnwolfe@google.com>
To: dev@dpdk.org, Junfeng Guo <junfeng.guo@intel.com>,
Joshua Washington <joshwash@google.com>
Cc: stable@dpdk.org, Harshitha Ramamurthy <hramamurthy@google.com>,
Rushil Gupta <rushilg@google.com>
Subject: [PATCH] net/gve: fix GQI QPL jumbo Tx FIFO reclaim and doorbell
Date: Sun, 15 Nov 2026 09:00:16 -0500 [thread overview]
Message-ID: <20261115140016.3830109-1-rnwolfe@google.com> (raw)
In GQI Queue Page List (GQI_QPL) mode each Tx queue copies outgoing
frames into a fixed QPL bounce FIFO (512 pages = 2 MiB on GCE, ~236
8,832-byte cacheline-aligned jumbo frames) even when nb_tx_desc is
configured to 2048. Under sustained jumbo bursts three bugs cause
unnecessary packet drops and mbuf tracking hazards:
1. Entry cleanup in gve_tx_burst_qpl() only ran when nb_free < free_thresh
(32) or fifo_avail == 0. With 2048 descriptors, 236 jumbo frames leave
nb_free = 1811 and non-zero remainder bytes in fifo_avail, so entry
cleanup never ran and every QPL wrap stalled mid-burst. Reclaim when
fifo_avail drops below half of fifo_size, and do not count normal
FIFO/ring-full backpressure as hardware stats.errors.
2. When is_fifo_avail() failed mid-burst after queuing packets, the driver
polled gve_tx_clean() once before writing qtx_tail, so the device had
not yet started draining the newly queued frames, and a single
GVE_TX_MAX_FREE_SZ batch after many small frames could stop short of
freeing a jumbo slot. Flush qtx_tail first and loop gve_tx_clean()
while next_to_clean advances.
3. sw_ring[sw_id] and the first TSO header slice were allocated/written
before verifying total FIFO space, gve_free_bulk_mbuf() did not clear
the first freed slot, and gve_tx_burst_ra() skipped mid-burst ring
cleanup and wrote pkt_len instead of data_len into the first segment
descriptor of multi-segment non-TSO packets.
Fixes: a46583cf43c8 ("net/gve: support Rx/Tx")
Cc: stable@dpdk.org
Signed-off-by: Ryan Wolfe <rnwolfe@google.com>
---
drivers/net/gve/gve_tx.c | 68 +++++++++++++++++++++++++++-------------
1 file changed, 47 insertions(+), 21 deletions(-)
diff --git a/drivers/net/gve/gve_tx.c b/drivers/net/gve/gve_tx.c
index 279669165f..2cf091f5be 100644
--- a/drivers/net/gve/gve_tx.c
+++ b/drivers/net/gve/gve_tx.c
@@ -19,6 +19,7 @@ gve_free_bulk_mbuf(struct rte_mbuf **txep, int num)
for (s = 0; s < num; s++) {
if (txep[s] != NULL) {
m = rte_pktmbuf_prefree_seg(txep[s]);
+ txep[s] = NULL;
if (m != NULL)
break;
}
@@ -200,7 +201,7 @@ gve_tx_fill_seg_desc(volatile union gve_tx_desc *desc, uint64_t ol_flags,
}
static inline bool
-is_fifo_avail(struct gve_tx_queue *txq, uint16_t len)
+is_fifo_avail(struct gve_tx_queue *txq, uint32_t len)
{
if (txq->fifo_avail < len)
return false;
@@ -210,6 +211,26 @@ is_fifo_avail(struct gve_tx_queue *txq, uint16_t len)
return false;
return true;
}
+
+static inline bool
+gve_tx_reclaim_fifo_qpl(struct gve_tx_queue *txq, uint32_t tx_tail,
+ uint32_t need)
+{
+ if (txq->tx_tail != tx_tail) {
+ rte_write32(rte_cpu_to_be_32(tx_tail), txq->qtx_tail);
+ txq->tx_tail = tx_tail;
+ }
+
+ do {
+ uint32_t prev_ntc = txq->next_to_clean;
+
+ gve_tx_clean(txq);
+ if (is_fifo_avail(txq, need))
+ return true;
+ if (txq->next_to_clean == prev_ntc)
+ return false;
+ } while (true);
+}
static inline uint64_t
gve_tx_alloc_from_fifo(struct gve_tx_queue *txq, uint16_t tx_id, uint16_t len)
{
@@ -266,7 +287,8 @@ gve_tx_burst_qpl(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
txr = txq->tx_desc_ring;
- if (txq->nb_free < txq->free_thresh || txq->fifo_avail == 0)
+ if (txq->nb_free < txq->free_thresh ||
+ txq->fifo_avail < (txq->fifo_size >> 1))
gve_tx_clean(txq);
if (txq->sw_nb_free < txq->free_thresh)
@@ -304,22 +326,23 @@ gve_tx_burst_qpl(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
(uint32_t)(tx_offload.l2_len + tx_offload.l3_len + tx_offload.l4_len) :
tx_pkt->pkt_len;
- if (!is_fifo_avail(txq, hlen)) {
- gve_tx_clean(txq);
- if (!is_fifo_avail(txq, hlen))
+ if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
+ uint32_t data_len, need;
+
+ if (unlikely(tx_pkt->pkt_len <= hlen))
+ goto end_of_tx;
+ data_len = tx_pkt->pkt_len - hlen;
+ need = RTE_ALIGN(hlen, RTE_CACHE_LINE_SIZE) +
+ RTE_ALIGN(data_len, RTE_CACHE_LINE_SIZE);
+ if (!is_fifo_avail(txq, need) &&
+ !gve_tx_reclaim_fifo_qpl(txq, tx_tail, need))
goto end_of_tx;
+ } else if (!is_fifo_avail(txq, hlen) &&
+ !gve_tx_reclaim_fifo_qpl(txq, tx_tail, hlen)) {
+ goto end_of_tx;
}
fifo_addr = gve_tx_alloc_from_fifo(txq, tx_id, hlen);
- /* For TSO, check if there's enough fifo space for data first */
- if (ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
- if (!is_fifo_avail(txq, tx_pkt->pkt_len - hlen)) {
- gve_tx_clean(txq);
- if (!is_fifo_avail(txq, tx_pkt->pkt_len - hlen))
- goto end_of_tx;
- }
- }
-
qpl_write_addr = (void *)(size_t)(fifo_addr + txq->fifo_base);
mbuf_header_addr = rte_pktmbuf_read(tx_pkt, 0, hlen, qpl_write_addr);
@@ -370,13 +393,14 @@ gve_tx_burst_qpl(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
end_of_tx:
if (nb_tx) {
- rte_write32(rte_cpu_to_be_32(tx_tail), txq->qtx_tail);
- txq->tx_tail = tx_tail;
+ if (txq->tx_tail != tx_tail) {
+ rte_write32(rte_cpu_to_be_32(tx_tail), txq->qtx_tail);
+ txq->tx_tail = tx_tail;
+ }
txq->sw_tail = sw_id;
txq->stats.packets += nb_tx;
txq->stats.bytes += bytes;
- txq->stats.errors += nb_pkts - nb_tx;
}
return nb_tx;
@@ -411,8 +435,11 @@ gve_tx_burst_ra(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
if (ol_flags & RTE_MBUF_F_TX_TCP_SEG)
nb_used++;
- if (txq->nb_free < nb_used)
- goto end_of_tx;
+ if (txq->nb_free < nb_used) {
+ gve_tx_clean(txq);
+ if (txq->nb_free < nb_used)
+ goto end_of_tx;
+ }
tx_offload.l2_len = tx_pkt->l2_len;
tx_offload.l3_len = tx_pkt->l3_len;
@@ -424,7 +451,7 @@ gve_tx_burst_ra(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
hlen = ol_flags & RTE_MBUF_F_TX_TCP_SEG ?
(uint32_t)(tx_offload.l2_len + tx_offload.l3_len + tx_offload.l4_len) :
- tx_pkt->pkt_len;
+ tx_pkt->data_len;
/*
* if tso, the driver needs to fill 2 descs for 1 mbuf
* so only put this mbuf into the 1st tx entry in sw ring
@@ -465,7 +492,6 @@ gve_tx_burst_ra(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
txq->stats.packets += nb_tx;
txq->stats.bytes += bytes;
- txq->stats.errors += nb_pkts - nb_tx;
}
return nb_tx;
--
2.56.0.679.gc527c65306-goog
next reply other threads:[~2026-11-17 7:58 UTC|newest]
Thread overview: 2+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-15 14:00 Ryan Wolfe [this message]
2026-11-17 17:34 ` [PATCH] net/gve: fix GQI QPL jumbo Tx FIFO reclaim and doorbell Stephen Hemminger
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20261115140016.3830109-1-rnwolfe@google.com \
--to=rnwolfe@google.com \
--cc=dev@dpdk.org \
--cc=hramamurthy@google.com \
--cc=joshwash@google.com \
--cc=junfeng.guo@intel.com \
--cc=rushilg@google.com \
--cc=stable@dpdk.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.