[net-next,21/24] net: marvell: Use nested-BH locking for XDP redirect.
Commit Message
The per-CPU variables used during bpf_prog_run_xdp() invocation and
later during xdp_do_redirect() rely on disabled BH for their protection.
Without locking in local_bh_disable() on PREEMPT_RT these data structure
require explicit locking.
This is a follow-up on the previous change which introduced
bpf_run_lock.redirect_lock and uses it now within drivers.
The simple way is to acquire the lock before bpf_prog_run_xdp() is
invoked and hold it until the end of function.
This does not always work because some drivers (cpsw, atlantic) invoke
xdp_do_flush() in the same context.
Acquiring the lock in bpf_prog_run_xdp() and dropping in
xdp_do_redirect() (without touching drivers) does not work because not
all driver, which use bpf_prog_run_xdp(), do support XDP_REDIRECT (and
invoke xdp_do_redirect()).
Ideally the minimal locking scope would be bpf_prog_run_xdp() +
xdp_do_redirect() and everything else (error recovery, DMA unmapping,
free/ alloc of memory, …) would happen outside of the locked section.
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Geetha sowjanya <gakula@marvell.com>
Cc: Jesper Dangaard Brouer <hawk@kernel.org>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: Marcin Wojtas <mw@semihalf.com>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Subbaraya Sundeep <sbhatta@marvell.com>
Cc: Sunil Goutham <sgoutham@marvell.com>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: hariprasad <hkelam@marvell.com>
Cc: bpf@vger.kernel.org
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
drivers/net/ethernet/marvell/mvneta.c | 2 ++
drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c | 1 +
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c | 1 +
3 files changed, 4 insertions(+)
@@ -2263,6 +2263,8 @@ mvneta_run_xdp(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
len = xdp->data_end - xdp->data_hard_start - pp->rx_offset_correction;
data_len = xdp->data_end - xdp->data;
+
+ guard(local_lock_nested_bh)(&bpf_run_lock.redirect_lock);
act = bpf_prog_run_xdp(prog, xdp);
/* Due xdp_adjust_tail: DMA sync for_device cover max len CPU touch */
@@ -3793,6 +3793,7 @@ mvpp2_run_xdp(struct mvpp2_port *port, struct bpf_prog *prog,
u32 ret, act;
len = xdp->data_end - xdp->data_hard_start - MVPP2_SKB_HEADROOM;
+ guard(local_lock_nested_bh)(&bpf_run_lock.redirect_lock);
act = bpf_prog_run_xdp(prog, xdp);
/* Due xdp_adjust_tail: DMA sync for_device cover max len CPU touch */
@@ -1422,6 +1422,7 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
xdp_prepare_buff(&xdp, hard_start, data - hard_start,
cqe->sg.seg_size, false);
+ guard(local_lock_nested_bh)(&bpf_run_lock.redirect_lock);
act = bpf_prog_run_xdp(prog, &xdp);
switch (act) {