Message-ID: <20200401160621.4fq66xwamuhmzxdb@soft-dev3.microsemi.net>
Date:   Wed, 1 Apr 2020 18:06:21 +0200
From:   Horatiu Vultur <horatiu.vultur@...rochip.com>
To:     Nikolay Aleksandrov <nikolay@...ulusnetworks.com>
CC:     <davem@...emloft.net>, <jiri@...nulli.us>, <ivecera@...hat.com>,
        <kuba@...nel.org>, <roopa@...ulusnetworks.com>,
        <olteanv@...il.com>, <andrew@...n.ch>,
        <UNGLinuxDriver@...rochip.com>, <linux-kernel@...r.kernel.org>,
        <netdev@...r.kernel.org>, <bridge@...ts.linux-foundation.org>
Subject: Re: [RFC net-next v4 7/9] bridge: mrp: Connect MRP api with the
 switchev API

The 03/30/2020 19:11, Nikolay Aleksandrov wrote:
> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
> 
> On 27/03/2020 11:21, Horatiu Vultur wrote:
> > Implement the MRP api.
> >
> > In case the HW can't generate MRP Test frames then the SW will try to generate
> > the frames. In case that also the SW will fail in generating the frames then a
> > error is return to the userspace. The userspace is responsible to generate all
> > the other MRP frames regardless if the test frames are generated by HW or SW.
> >
> > The forwarding/termination of MRP frames is happening in the kernel and is done
> > by the MRP instance. The userspace application doesn't do the forwarding.
> >
> > Signed-off-by: Horatiu Vultur <horatiu.vultur@...rochip.com>
> > ---
> >  net/bridge/br_mrp.c | 514 ++++++++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 514 insertions(+)
> >  create mode 100644 net/bridge/br_mrp.c
> >
> 
> Hi,

Hi Nik,

> In general the RCU usage needs more work.

Thanks for the detailed review, this is my first time when I use the RCU,
so I might need to spend more time on time.

> Also I might've missed it, but where do you
> handle bridge port delete which is used in mrp ?

When a port is deleted, then the userspace application will be notified
and then the userspace will remove the MRP instance. Because there is no
point to have a MRP instance with only 1 port. And the function that
delets the MRP instance is br_mrp_del.

> Also do you actually need the mrp->lock ?

I think I should be fine not to use mrp->lock because already the rtnl
lock is taken.

> 
> > diff --git a/net/bridge/br_mrp.c b/net/bridge/br_mrp.c
> > new file mode 100644
> > index 000000000000..f1de792d7a6e
> > --- /dev/null
> > +++ b/net/bridge/br_mrp.c
> > @@ -0,0 +1,514 @@
> > +// SPDX-License-Identifier: GPL-2.0-or-later
> > +
> > +#include "br_private_mrp.h"
> > +
> > +static const u8 mrp_test_dmac[ETH_ALEN] = { 0x1, 0x15, 0x4e, 0x0, 0x0, 0x1 };
> > +
> > +static struct net_bridge_port *br_mrp_get_port(struct net_bridge *br,
> > +                                            u32 ifindex)
> > +{
> > +     struct net_bridge_port *res = NULL;
> > +     struct net_bridge_port *port;
> > +
> > +     spin_lock_bh(&br->lock);
> > +
> 
> This is called under RTNL, you don't need the br->lock.

Will be fix in the next patch series.
> 
> > +     list_for_each_entry(port, &br->port_list, list) {
> > +             if (port->dev->ifindex == ifindex) {
> > +                     res = port;
> > +                     break;
> > +             }
> > +     }
> > +
> > +     spin_unlock_bh(&br->lock);
> > +
> > +     return res;
> > +}
> > +
> > +static struct br_mrp *br_mrp_find_id(struct net_bridge *br, u32 ring_id)
> > +{
> > +     struct br_mrp *res = NULL;
> > +     struct br_mrp *mrp;
> > +
> > +     rcu_read_lock();
> > +
> 
> This is generally a bad pattern because it can hide legitimate bugs and make
> it harder to debug.

Can you give me a little more details why is a bad pattern?
I have tried to read about rcu from here[1][2]. But I couldn't see
anything about this.

> 
> > +     list_for_each_entry_rcu(mrp, &br->mrp_list, list) {
> > +             if (mrp->ring_id == ring_id) {
> > +                     res = mrp;
> > +                     break;
> > +             }
> > +     }
> > +
> > +     rcu_read_unlock();
> > +
> > +     return res;
> > +}
> > +
> > +static struct br_mrp *br_mrp_find_port(struct net_bridge *br,
> > +                                    struct net_bridge_port *p)
> > +{
> > +     struct br_mrp *res = NULL;
> > +     struct br_mrp *mrp;
> > +
> > +     rcu_read_lock();
> > +
> > +     list_for_each_entry_rcu(mrp, &br->mrp_list, list) {
> > +             if (rcu_dereference(mrp->p_port) == p ||
> > +                 rcu_dereference(mrp->s_port) == p) {
> 
> rcu_access_pointer() is ok for comparisons

Will be fix in the next patch series.
> 
> > +                     res = mrp;
> > +                     break;
> > +             }
> > +     }
> > +
> > +     rcu_read_unlock();
> > +
> > +     return res;
> > +}
> > +
> > +static int br_mrp_next_seq(struct br_mrp *mrp)
> > +{
> > +     mrp->seq_id++;
> > +     return mrp->seq_id;
> > +}
> > +
> > +static struct sk_buff *br_mrp_skb_alloc(struct net_bridge_port *p,
> > +                                     const u8 *src, const u8 *dst)
> > +{
> > +     struct ethhdr *eth_hdr;
> > +     struct sk_buff *skb;
> > +     u16 *version;
> > +
> > +     skb = dev_alloc_skb(MRP_MAX_FRAME_LENGTH);
> > +     if (!skb)
> > +             return NULL;
> > +
> > +     skb->dev = p->dev;
> > +     skb->protocol = htons(ETH_P_MRP);
> > +     skb->priority = MRP_FRAME_PRIO;
> > +     skb_reserve(skb, sizeof(*eth_hdr));
> > +
> > +     eth_hdr = skb_push(skb, sizeof(*eth_hdr));
> > +     ether_addr_copy(eth_hdr->h_dest, dst);
> > +     ether_addr_copy(eth_hdr->h_source, src);
> > +     eth_hdr->h_proto = htons(ETH_P_MRP);
> > +
> > +     version = skb_put(skb, sizeof(*version));
> > +     *version = cpu_to_be16(MRP_VERSION);
> > +
> > +     return skb;
> > +}
> > +
> > +static void br_mrp_skb_tlv(struct sk_buff *skb,
> > +                        enum br_mrp_tlv_header_type type,
> > +                        u8 length)
> > +{
> > +     struct br_mrp_tlv_hdr *hdr;
> > +
> > +     hdr = skb_put(skb, sizeof(*hdr));
> > +     hdr->type = type;
> > +     hdr->length = length;
> > +}
> > +
> > +static void br_mrp_skb_common(struct sk_buff *skb, struct br_mrp *mrp)
> > +{
> > +     struct br_mrp_common_hdr *hdr;
> > +
> > +     br_mrp_skb_tlv(skb, BR_MRP_TLV_HEADER_COMMON, sizeof(*hdr));
> > +
> > +     hdr = skb_put(skb, sizeof(*hdr));
> > +     hdr->seq_id = cpu_to_be16(br_mrp_next_seq(mrp));
> > +     memset(hdr->domain, 0xff, MRP_DOMAIN_UUID_LENGTH);
> > +}
> > +
> > +static struct sk_buff *br_mrp_alloc_test_skb(struct br_mrp *mrp,
> > +                                          struct net_device *dev,
> > +                                          enum br_mrp_port_role_type port_role)
> > +{
> > +     struct net_bridge_port *p = br_port_get_rtnl(dev);
> > +     struct br_mrp_ring_test_hdr *hdr = NULL;
> > +     struct net_bridge *br = p->br;
> > +     struct sk_buff *skb = NULL;
> > +
> > +     if (!p)
> > +             return NULL;
> > +
> > +     br = p->br;
> > +
> > +     skb = br_mrp_skb_alloc(p, p->dev->dev_addr, mrp_test_dmac);
> > +     if (!skb)
> > +             return NULL;
> > +
> > +     br_mrp_skb_tlv(skb, BR_MRP_TLV_HEADER_RING_TEST, sizeof(*hdr));
> > +     hdr = skb_put(skb, sizeof(*hdr));
> > +
> > +     hdr->prio = cpu_to_be16(MRP_DEFAULT_PRIO);
> > +     ether_addr_copy(hdr->sa, p->br->dev->dev_addr);
> > +     hdr->port_role = cpu_to_be16(port_role);
> > +     hdr->state = cpu_to_be16(mrp->ring_state);
> > +     hdr->transitions = cpu_to_be16(mrp->ring_transitions);
> > +     hdr->timestamp = cpu_to_be32(jiffies_to_msecs(jiffies));
> > +
> > +     br_mrp_skb_common(skb, mrp);
> > +     br_mrp_skb_tlv(skb, BR_MRP_TLV_HEADER_END, 0x0);
> > +
> > +     return skb;
> > +}
> > +
> > +static void br_mrp_test_work_expired(struct work_struct *work)
> > +{
> > +     struct delayed_work *del_work = to_delayed_work(work);
> > +     struct br_mrp *mrp = container_of(del_work, struct br_mrp, test_work);
> > +     bool notify_open = false;
> > +     struct sk_buff *skb;
> > +
> 
> Since this runs asynchronously what happens if the port is deleted ?

Later I have checks to see if the port is no NULL. Is not good enough?
I have these rcu_access_pointer checks and before that I disable the
interrupts and get the rcu lock.

> 
> > +     if (time_before_eq(mrp->test_end, jiffies))
> > +             return;
> > +
> > +     if (mrp->test_count_miss < mrp->test_max_miss) {
> > +             mrp->test_count_miss++;
> > +     } else {
> > +             /* Notify that the ring is open only if the ring state is
> > +              * closed, otherwise it would continue to notify at every
> > +              * interval.
> > +              */
> > +             if (mrp->ring_state == BR_MRP_RING_STATE_CLOSED)
> > +                     notify_open = true;
> > +     }
> > +
> > +     local_bh_disable();
> > +     rcu_read_lock();
> > +
> > +     if (!rcu_access_pointer(mrp->p_port) ||
> > +         !rcu_access_pointer(mrp->s_port))
> > +             goto out;
> > +
> > +     /* Is it possible here to get call to delete the bridge port? If yes
> > +      * I need to protect it
> > +      */
> > +     dev_hold(rcu_dereference(mrp->p_port)->dev);
> > +
> 
> This looks all wrong, p_port can become NULL here and you'll deref it.

By disabling the interrupts and taking the rcu read lock, will I not be
sure that no one can access the p_port?
If is not true, how the p_port can become NULL?

> 
> > +     skb = br_mrp_alloc_test_skb(mrp, rcu_dereference(mrp->p_port)->dev,
> > +                                 BR_MRP_PORT_ROLE_PRIMARY);
> > +     if (!skb)
> > +             goto out;
> > +
> > +     skb_reset_network_header(skb);
> > +     dev_queue_xmit(skb);
> > +
> > +     if (notify_open && !mrp->ring_role_offloaded)
> > +             br_mrp_port_open(rcu_dereference(mrp->p_port)->dev, true);
> > +
> > +     dev_put(rcu_dereference(mrp->p_port)->dev);
> > +
> > +     dev_hold(rcu_dereference(mrp->s_port)->dev);
> > +
> 
> same here
> 
> > +     skb = br_mrp_alloc_test_skb(mrp, rcu_dereference(mrp->s_port)->dev,
> > +                                 BR_MRP_PORT_ROLE_SECONDARY);
> > +     if (!skb)
> > +             goto out;
> > +
> > +     skb_reset_network_header(skb);
> > +     dev_queue_xmit(skb);
> > +
> > +     if (notify_open && !mrp->ring_role_offloaded)
> > +             br_mrp_port_open(rcu_dereference(mrp->s_port)->dev, true);
> > +
> > +     dev_put(rcu_dereference(mrp->s_port)->dev);
> > +
> > +out:
> > +     rcu_read_unlock();
> > +     local_bh_enable();
> > +
> > +     queue_delayed_work(system_wq, &mrp->test_work,
> > +                        usecs_to_jiffies(mrp->test_interval));
> > +}
> > +
> > +/* Adds a new MRP instance.
> > + * note: called under rtnl_lock
> > + */
> > +int br_mrp_add(struct net_bridge *br, struct br_mrp_instance *instance)
> > +{
> > +     struct net_bridge_port *p;
> > +     struct br_mrp *mrp;
> > +
> > +     /* If the ring exists, it is not possible to create another one with the
> > +      * same ring_id
> > +      */
> > +     mrp = br_mrp_find_id(br, instance->ring_id);
> > +     if (mrp)
> > +             return -EINVAL;
> > +
> > +     if (!br_mrp_get_port(br, instance->p_ifindex) ||
> > +         !br_mrp_get_port(br, instance->s_ifindex))
> > +             return -EINVAL;
> > +
> > +     mrp = devm_kzalloc(&br->dev->dev, sizeof(struct br_mrp), GFP_KERNEL);
> > +     if (!mrp)
> > +             return -ENOMEM;
> > +
> > +     /* I think is not needed because this can be replaced with rtnl lock*/
> > +     spin_lock_init(&mrp->lock);
> > +     spin_lock(&mrp->lock);
> > +
> > +     mrp->br = br;
> 
> Is this field (mrp->br) used anywhere ?

Not anymore. I can remove it in the next patch series.

> 
> > +     mrp->ring_id = instance->ring_id;
> > +
> > +     p = br_mrp_get_port(br, instance->p_ifindex);
> > +     p->state = BR_STATE_FORWARDING;
> > +     p->flags |= BR_MRP_AWARE;
> > +     rcu_assign_pointer(mrp->p_port, p);
> > +
> > +     p = br_mrp_get_port(br, instance->s_ifindex);
> > +     p->state = BR_STATE_FORWARDING;
> > +     p->flags |= BR_MRP_AWARE;
> > +     rcu_assign_pointer(mrp->s_port, p);
> > +
> > +     br_mrp_switchdev_add(mrp);
> > +
> > +     spin_unlock(&mrp->lock);
> > +     synchronize_rcu();
> 
> Why do you need the synchronize here?

Actually this shouldn't be after the list_add_tail_rcu? Because I am
thinking that some can read the list at the same time I am change it.

> 
> > +
> > +     list_add_tail_rcu(&mrp->list, &br->mrp_list);
> > +     INIT_DELAYED_WORK(&mrp->test_work, br_mrp_test_work_expired);
> > +
> > +     return 0;
> > +}
> > +
> > +/* Deletes existing MRP instance.
> > + * note: called under rtnl_lock
> > + */
> > +int br_mrp_del(struct net_bridge *br, struct br_mrp_instance *instance)
> > +{
> > +     struct br_mrp *mrp = br_mrp_find_id(br, instance->ring_id);
> > +     struct net_bridge_port *p;
> > +
> > +     if (!mrp)
> > +             return -EINVAL;
> > +
> > +     /* Stop sending MRP_Test frames */
> > +     cancel_delayed_work(&mrp->test_work);
> 
> cancel_delayed_work_sync() if you'd like to make sure it's stopped and finished (if it was running
> during this)

Will be fixed in the next patch series.
> 
> > +     br_mrp_switchdev_send_ring_test(mrp, 0, 0, 0);
> > +
> > +     spin_lock(&mrp->lock);
> > +
> > +     br_mrp_switchdev_del(mrp);
> > +
> > +     /* Reset the ports */
> > +     p = rcu_dereference_protected(mrp->p_port, lockdep_is_held(&mrp->lock));
> > +     if (p) {
> > +             spin_lock(&br->lock);
> > +             p->state = BR_STATE_FORWARDING;
> > +             p->flags &= ~BR_MRP_AWARE;
> > +             br_mrp_port_switchdev_set_state(p, BR_STATE_FORWARDING);
> > +             rcu_assign_pointer(mrp->p_port, NULL);
> > +             spin_unlock(&br->lock);
> > +     }
> > +
> > +     p = rcu_dereference_protected(mrp->s_port, lockdep_is_held(&mrp->lock));
> > +     if (p) {
> > +             spin_lock(&br->lock);
> > +             p->state = BR_STATE_FORWARDING;
> > +             p->flags &= ~BR_MRP_AWARE;
> > +             br_mrp_port_switchdev_set_state(p, BR_STATE_FORWARDING);
> > +             rcu_assign_pointer(mrp->s_port, NULL);
> > +             spin_unlock(&br->lock);
> > +     }
> > +
> > +     /* Destroy the ring */
> > +     mrp->br = NULL;
> > +
> > +     spin_unlock(&mrp->lock);
> > +     synchronize_rcu();
> > +
> > +     list_del_rcu(&mrp->list);
> > +     devm_kfree(&br->dev->dev, mrp);
> > +
> > +     return 0;
> > +}
> > +
> > +int br_mrp_set_port_state(struct net_bridge_port *p,
> > +                       enum br_mrp_port_state_type state)
> > +{
> > +     spin_lock(&p->br->lock);
> > +
> > +     if (state == BR_MRP_PORT_STATE_FORWARDING)
> > +             p->state = BR_STATE_FORWARDING;
> > +     else
> > +             p->state = BR_STATE_BLOCKING;
> > +
> > +     br_mrp_port_switchdev_set_state(p, state);
> > +
> > +     spin_unlock(&p->br->lock);
> > +
> > +     return 0;
> > +}
> > +
> > +int br_mrp_set_port_role(struct net_bridge_port *p,
> > +                      u32 ring_id, enum br_mrp_port_role_type role)
> > +{
> > +     struct br_mrp *mrp = br_mrp_find_id(p->br, ring_id);
> > +
> > +     if (!mrp)
> > +             return -EINVAL;
> > +
> > +     spin_lock(&mrp->lock);
> > +
> > +     if (role == BR_MRP_PORT_ROLE_PRIMARY)
> > +             rcu_assign_pointer(mrp->p_port, p);
> > +     if (role == BR_MRP_PORT_ROLE_SECONDARY)
> > +             rcu_assign_pointer(mrp->s_port, p);
> > +
> > +     br_mrp_port_switchdev_set_role(p, role);
> > +
> > +     spin_unlock(&mrp->lock);
> > +     synchronize_rcu();
> 
> Why do you need to synchronize here?

Actually this is not needed.
> 
> > +
> > +     return 0;
> > +}
> > +
> > +int br_mrp_set_ring_state(struct net_bridge *br, u32 ring_id,
> > +                       enum br_mrp_ring_state_type state)
> > +{
> > +     struct br_mrp *mrp = br_mrp_find_id(br, ring_id);
> > +
> > +     if (!mrp)
> > +             return -EINVAL;
> > +
> > +     if (mrp->ring_state == BR_MRP_RING_STATE_CLOSED &&
> > +         state != BR_MRP_RING_STATE_CLOSED)
> > +             mrp->ring_transitions++;
> > +
> > +     mrp->ring_state = state;
> > +
> > +     br_mrp_switchdev_set_ring_state(mrp, state);
> > +
> > +     return 0;
> > +}
> > +
> > +int br_mrp_set_ring_role(struct net_bridge *br, u32 ring_id,
> > +                      enum br_mrp_ring_role_type role)
> > +{
> > +     struct br_mrp *mrp = br_mrp_find_id(br, ring_id);
> > +     int err;
> > +
> > +     if (!mrp)
> > +             return -EINVAL;
> > +
> > +     mrp->ring_role = role;
> > +
> > +     /* If there is an error just bailed out */
> > +     err = br_mrp_switchdev_set_ring_role(mrp, role);
> > +     if (err && err != -EOPNOTSUPP)
> > +             return err;
> > +
> > +     /* Now detect if the HW actually applied the role or not. If the HW
> > +      * applied the role it means that the SW will not to do those operations
> > +      * anymore. For example if the role ir MRM then the HW will notify the
> > +      * SW when ring is open, but if the is not pushed to the HW the SW will
> > +      * need to detect when the ring is open
> > +      */
> > +     mrp->ring_role_offloaded = err == -EOPNOTSUPP ? 0 : 1;
> > +
> > +     return 0;
> > +}
> > +
> > +int br_mrp_start_test(struct net_bridge *br, u32 ring_id, u32 interval,
> > +                   u8 max_miss, u32 period)
> > +{
> > +     struct br_mrp *mrp = br_mrp_find_id(br, ring_id);
> > +
> > +     if (!mrp)
> > +             return -EINVAL;
> > +
> > +     /* Try to push is to the HW and if it fails then continue to generate in
> > +      * SW and if that also fails then return error
> > +      */
> > +     if (!br_mrp_switchdev_send_ring_test(mrp, interval, max_miss, period))
> > +             return 0;
> > +
> > +     mrp->test_interval = interval;
> > +     mrp->test_end = jiffies + usecs_to_jiffies(period);
> > +     mrp->test_max_miss = max_miss;
> > +     mrp->test_count_miss = 0;
> > +     queue_delayed_work(system_wq, &mrp->test_work,
> > +                        usecs_to_jiffies(interval));
> > +
> > +     return 0;
> > +}
> > +
> > +/* Process only MRP Test frame. All the other MRP frames are processed by
> > + * userspace application
> > + * note: already called with rcu_read_lock
> > + */
> > +static void br_mrp_mrm_process(struct br_mrp *mrp, struct sk_buff *skb)
> > +{
> > +     struct br_mrp_tlv_hdr *hdr;
> > +
> > +     hdr = (struct br_mrp_tlv_hdr *)(skb->data + sizeof(uint16_t));
> > +
> > +     if (hdr->type != BR_MRP_TLV_HEADER_RING_TEST)
> > +             return;
> > +
> > +     mrp->test_count_miss = 0;
> > +
> > +     br_mrp_port_open(rcu_dereference(mrp->p_port)->dev, false);
> > +     br_mrp_port_open(rcu_dereference(mrp->s_port)->dev, false);
> > +}
> > +
> > +/* This will just forward the frame to the other mrp ring port(MRC role) or will
> > + * not do anything.
> > + * note: already called with rcu_read_lock
> > + */
> > +static int br_mrp_rcv(struct net_bridge_port *p,
> > +                   struct sk_buff *skb, struct net_device *dev)
> > +{
> > +     struct net_device *s_dev, *p_dev, *d_dev;
> > +     struct net_bridge *br;
> > +     struct sk_buff *nskb;
> > +     struct br_mrp *mrp;
> > +
> > +     /* If port is disable don't accept any frames */
> > +     if (p->state == BR_STATE_DISABLED)
> > +             return 0;
> > +
> > +     br = p->br;
> > +     mrp =  br_mrp_find_port(br, p);
> > +     if (unlikely(!mrp))
> > +             return 0;
> > +
> > +     /* If the role is MRM then don't forward the frames */
> > +     if (mrp->ring_role == BR_MRP_RING_ROLE_MRM) {
> > +             br_mrp_mrm_process(mrp, skb);
> > +             return 1;
> > +     }
> > +
> > +     nskb = skb_clone(skb, GFP_ATOMIC);
> > +     if (!nskb)
> > +             return 0;
> > +
> > +     p_dev = rcu_dereference(mrp->p_port)->dev;
> > +     s_dev = rcu_dereference(mrp->s_port)->dev;
> > +
> 
> Not safe, could deref null.

Will be fixed in the next patch series.

> 
> > +     if (p_dev == dev)
> > +             d_dev = s_dev;
> > +     else
> > +             d_dev = p_dev;
> > +
> > +     nskb->dev = d_dev;
> > +     skb_push(nskb, ETH_HLEN);
> > +     dev_queue_xmit(nskb);
> > +
> > +     return 1;
> > +}
> > +
> > +int br_mrp_process(struct net_bridge_port *p, struct sk_buff *skb)
> > +{
> > +     /* If there is no MRP instance do normal forwarding */
> > +     if (unlikely(!(p->flags & BR_MRP_AWARE)))
> 
> Shouldn't this one be likely() ?

Yes, this should be likely.
> 
> > +             goto out;
> > +
> > +     if (unlikely(skb->protocol == htons(ETH_P_MRP)))
> > +             return br_mrp_rcv(p, skb, p->dev);
> > +
> > +out:
> > +     return 0;
> > +}
> >
> 

[1] https://lwn.net/Articles/262464/
[2] https://www.kernel.org/doc/html/latest/RCU/listRCU.html

-- 
/Horatiu

Powered by blists - more mailing lists