Kyle Zeng wrote:
AF_PACKET TX_RING skbs keep a raw pointer to their ring frame. The skb
page references preserve page-backed ring blocks after pg_vec is freed,
but they do not preserve a vmalloc mapping.
Claude shows an interesting case where this page-backed statement does
not hold: if the entire skb is linear. Not for this patch, but a
similar case.
tpacket_destruct_skb() currently drops the pending reference before
writing the timestamp and TP_STATUS_AVAILABLE to the frame. Move the
decrement after those stores. The smp_wmb() in __packet_set_status()
orders the frame stores before the decrement.
Also recheck pending TX frames under pg_vec_lock before non-closing
ring replacement, so a racing send cannot add a pending skb between
the initial check and the ring swap.
Ring allocation can produce a mixture of page-backed and vmalloc-backed
blocks. Allocate deferred-work storage during TX ring setup when the
first vmalloc-backed block is encountered, and keep its pointer in the
pg_vec allocation header. If allocation fails, return -ENOMEM from ring
setup. On socket close, a non-NULL pointer identifies a vmalloc-backed
vector without a scan. If TX skbs remain, defer the whole vector to
system_long_wq.
After pg_vec is detached, a late destructor can skip the pending
decrement. Use socket write-memory accounting as the deferred lifetime
gate instead: an skb remains charged through its final sock_wfree(),
after all ring-frame accesses. The delayed work retains a socket
reference and reschedules itself until no TX skbs remain.
Move pending_refcnt release to packet_sock_destruct() so late skb
destructors and deferred cleanup can safely use it after
packet_release(). Page-backed teardown remains synchronous, and no lock
is added to the TX completion hot path.
This one patch combines multiple fixes. If and only if a respin is
needed, it may be good to break it up to help understanding.
Fixes: b013840810c2 ("packet: use percpu mmap tx frame pending refcount")
Cc: <redacted>
Link: https://lore.kernel.org/netdev/20260721015824.45829-1-kylebot@openai.com/ (local)
Suggested-by: Eric Dumazet <redacted>
Suggested-by: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Assisted-by: Codex:gpt-5.6-sol
quoted hunk ↗ jump to hunk
+struct packet_pg_vec {
+ struct packet_pg_vec_free *deferred;
+ unsigned int order;
+ unsigned int len;
+ struct pgv pg_vec[] __counted_by(len);
+};
+
+struct packet_pg_vec_free {
+ struct delayed_work work;
+ struct sock *sk;
+ struct packet_pg_vec *vec;
+};
quoted hunk ↗ jump to hunk
@@ -4382,7 +4399,46 @@ static void free_pg_vec(struct pgv *pg_vec, unsigned int order,
+static void packet_free_pg_vec_work(struct work_struct *work)
+{
+ struct packet_pg_vec_free *deferred;
+ struct packet_pg_vec *vec;
+ struct sock *sk;
+
+ deferred = container_of_const(to_delayed_work(work),
+ struct packet_pg_vec_free, work);
+ vec = deferred->vec;
+ sk = deferred->sk;
+ if (sk_wmem_alloc_get(sk)) {
+ queue_delayed_work(system_long_wq, &deferred->work, 1);
Can this keep requeueing itself? Claude suggests using pending ring
count as gate.
quoted hunk ↗ jump to hunk
+static void packet_free_tx_ring(struct sock *sk, struct pgv *pg_vec,
+ unsigned int order, unsigned int len)
+{
+ struct packet_pg_vec_free *deferred;
+ struct packet_pg_vec *vec;
+
+ vec = container_of_const(pg_vec, struct packet_pg_vec, pg_vec[0]);
+ deferred = vec->deferred;
+ if (!deferred || !sk_wmem_alloc_get(sk)) {
+ free_pg_vec(pg_vec, order, len);
+ return;
+ }
+
+ /* A detached ring's pending count can miss late skb destructors. */
+ deferred->sk = sk;
+ sock_hold(sk);
+ queue_delayed_work(system_long_wq, &deferred->work, 0);
}
static char *alloc_one_pg_vec_page(unsigned long order)@@ -4410,20 +4466,35 @@ static char *alloc_one_pg_vec_page(unsigned long order)
return NULL;
}
-static struct pgv *alloc_pg_vec(struct tpacket_req *req, int order)
+static struct pgv *alloc_pg_vec(struct tpacket_req *req, int order, bool tx_ring)
{
unsigned int block_nr = req->tp_block_nr;
+ struct packet_pg_vec *vec;
struct pgv *pg_vec;
int i;
- pg_vec = kzalloc_objs(struct pgv, block_nr, GFP_KERNEL | __GFP_NOWARN);
- if (unlikely(!pg_vec))
- goto out;
+ vec = kzalloc_flex(*vec, pg_vec, block_nr, GFP_KERNEL | __GFP_NOWARN);
+ if (unlikely(!vec))
+ return NULL;
+ vec->order = order;
+ vec->len = block_nr;
+ pg_vec = vec->pg_vec;
for (i = 0; i < block_nr; i++) {
pg_vec[i].buffer = alloc_one_pg_vec_page(order);
if (unlikely(!pg_vec[i].buffer))
goto out_free_pgvec;
+
+ if (tx_ring && !vec->deferred &&
+ is_vmalloc_addr(pg_vec[i].buffer)) {
+ vec->deferred = kzalloc_obj(*vec->deferred,
+ GFP_KERNEL | __GFP_NOWARN);
+ if (!vec->deferred)
+ goto out_free_pgvec;
+ vec->deferred->vec = vec;
The nested structures are fairly complex.
Would it make sense to avoid the separate packet_pg_vec_free, fold
that into packet_pg_vec and use a different field to identify
whether vmalloc backed pages are used. Maybe sk, or even a new
boolean field has_vmalloc, for readability