On Thu, Oct 11, 2018 at 01:11:59AM +0200, Frederic Weisbecker wrote:
> This pair of function is implemented on top of local_bh_disable() that
> is going to handle a softirq mask in order to apply finegrained vector
> disablement. The lock function is going to return the previous vectors
> enabled mask prior to the last call to local_bh_disable(), following a
> similar model to that of local_irq_save/restore. Subsequent calls to
> local_bh_disable() and friends can then stack up:
>
> bh = local_bh_disable(vec_mask);
> bh2 = rcu_read_lock_bh() {
> bh2 = local_bh_disable(...)
> return bh2;
> }
> ...
> rcu_read_unlock_bh(bh2) {
> local_bh_enable(bh2);
> }
> local_bh_enable(bh);
>
> To prepare for that, make rcu_read_lock_bh() able to return a saved vector
> enabled mask and pass it back to rcu_read_unlock_bh(). We'll plug it
> to local_bh_disable() in a subsequent patch.
> Signed-off-by: Frederic Weisbecker <frede...@kernel.org>
> Cc: Ingo Molnar <mi...@kernel.org>
> Cc: Sebastian Andrzej Siewior <bige...@linutronix.de>
> Cc: Thomas Gleixner <t...@linutronix.de>
> Cc: Peter Zijlstra <pet...@infradead.org>
> Cc: Linus Torvalds <torva...@linux-foundation.org>
> Cc: David S. Miller <da...@davemloft.net>
> Cc: Mauro Carvalho Chehab <mche...@s-opensource.com>
> Cc: Paul E. McKenney <paul...@linux.vnet.ibm.com>
> ---
> crypto/pcrypt.c | 5 ++--
> drivers/infiniband/ulp/ipoib/ipoib_main.c | 5 ++--
> drivers/net/hyperv/rndis_filter.c | 5 ++--
> drivers/net/macsec.c | 12 +++++----
> drivers/net/vrf.c | 19 ++++++++------
> drivers/vhost/net.c | 5 ++--
> include/linux/rcupdate.h | 5 ++--
> include/net/arp.h | 10 ++++---
> include/net/ip6_fib.h | 1 +
> include/net/ndisc.h | 10 ++++---
> include/net/neighbour.h | 1 +
> kernel/padata.c | 5 ++--
> kernel/rcu/rcuperf.c | 2 +-
> kernel/rcu/rcutorture.c | 2 +-
> net/caif/caif_dev.c | 5 ++--
> net/core/dev.c | 7 ++---
> net/core/neighbour.c | 37 +++++++++++++++-----------
> net/core/pktgen.c | 5 ++--
> net/decnet/dn_route.c | 27 +++++++++++--------
> net/ipv4/fib_semantics.c | 5 ++--
> net/ipv4/ip_output.c | 7 ++---
> net/ipv4/netfilter/ipt_CLUSTERIP.c | 5 ++--
> net/ipv6/addrconf.c | 21 ++++++++-------
> net/ipv6/ip6_fib.c | 4 +--
> net/ipv6/ip6_flowlabel.c | 43
> ++++++++++++++++++-------------
> net/ipv6/ip6_output.c | 12 +++++----
> net/ipv6/route.c | 15 ++++++-----
> net/ipv6/xfrm6_tunnel.c | 5 ++--
> net/l2tp/l2tp_core.c | 33 ++++++++++++++----------
> net/llc/llc_core.c | 5 ++--
> net/llc/llc_proc.c | 13 +++++++---
> net/llc/llc_sap.c | 5 ++--
> net/netfilter/ipset/ip_set_core.c | 10 ++++---
> net/netfilter/ipset/ip_set_hash_gen.h | 15 ++++++-----
> net/netfilter/nfnetlink_log.c | 17 ++++++++----
> 35 files changed, 229 insertions(+), 154 deletions(-)
>
> diff --git a/crypto/pcrypt.c b/crypto/pcrypt.c
> index f8ec3d4..490358c 100644
> --- a/crypto/pcrypt.c
> +++ b/crypto/pcrypt.c
> @@ -73,12 +73,13 @@ struct pcrypt_aead_ctx {
> static int pcrypt_do_parallel(struct padata_priv *padata, unsigned int
> *cb_cpu,
> struct padata_pcrypt *pcrypt)
> {
> + unsigned int bh;
> unsigned int cpu_index, cpu, i;
> struct pcrypt_cpumask *cpumask;
>
> cpu = *cb_cpu;
>
> - rcu_read_lock_bh();
> + bh = rcu_read_lock_bh();
> cpumask = rcu_dereference_bh(pcrypt->cb_cpumask);
> if (cpumask_test_cpu(cpu, cpumask->mask))
> goto out;
> @@ -95,7 +96,7 @@ static int pcrypt_do_parallel(struct padata_priv *padata,
> unsigned int *cb_cpu,
> *cb_cpu = cpu;
>
> out:
> - rcu_read_unlock_bh();
> + rcu_read_unlock_bh(bh);
> return padata_do_parallel(pcrypt->pinst, padata, cpu);
> }
This complicates the RCU API for -bh and doesn't look pretty at all. Is there
anything better we can do so we don't have to touch existing readers at all?
Also, I thought softirqs were kind of a thing of the past, and threaded
interrupts are the more preferred interrupt bottom halves these days,
especially for -rt. Maybe that was just wishful thinking on my part :-)
thanks,
- Joel
