Date:   Fri, 26 Mar 2021 01:24:58 +0900
From:   Taehee Yoo <ap420073@...il.com>
To:     netdev@...r.kernel.org, davem@...emloft.net, kuba@...nel.org
Cc:     jwi@...ux.ibm.com, kgraul@...ux.ibm.com, hca@...ux.ibm.com,
        gor@...ux.ibm.com, borntraeger@...ibm.com,
        mareklindner@...mailbox.ch, sw@...onwunderlich.de, a@...table.cc,
        sven@...fation.org, yoshfuji@...ux-ipv6.org, dsahern@...nel.org,
        linux-s390@...r.kernel.org, b.a.t.m.a.n@...ts.open-mesh.org,
        Cong Wang <xiyou.wangcong@...il.com>
Subject: Re: [PATCH net-next v3 0/7] mld: change context from atomic to
 sleepable


On 3/26/21 1:16 AM, Taehee Yoo wrote:
> This patchset changes the context of MLD module.
> Before this patchset, MLD functions are atomic context so it couldn't use
> sleepable functions and flags.
> 
> There are several reasons why MLD functions are under atomic context.
> 1. It uses timer API.
> Timer expiration functions are executed in the atomic context.
> 2. atomic locks
> MLD functions use rwlock and spinlock to protect their own resources.
> 
> So, in order to switch context, this patchset converts resources to use
> RCU and removes atomic locks and timer API.
> 
> 1. The first patch convert from the timer API to delayed work.
> Timer API is used for delaying some works.
> MLD protocol has a delay mechanism, which is used for replying to a query.
> If a listener receives a query from a router, it should send a response
> after some delay. But because of timer expire function is executed in
> the atomic context, this patch convert from timer API to the delayed work.
> 
> 2. The fourth patch deletes inet6_dev->mc_lock.
> The mc_lock has protected inet6_dev->mc_tomb pointer.
> But this pointer is already protected by RTNL and it isn't be used by
> datapath. So, it isn't be needed and because of this, many atomic context
> critical sections are deleted.
> 
> 3. The fifth patch convert ip6_sf_socklist to RCU.
> ip6_sf_socklist has been protected by ipv6_mc_socklist->sflock(rwlock).
> But this is already protected by RTNL So if it is converted to use RCU
> in order to be used in the datapath, the sflock is no more needed.
> So, its control path context can be switched to sleepable.
> 
> 4. The sixth patch convert ip6_sf_list to RCU.
> The reason for this patch is the same as the previous patch.
> 
> 5. The seventh patch convert ifmcaddr6 to RCU.
> The reason for this patch is the same as the previous patch.
> 
> 6. Add new workqueues for processing query/report event.
> By this patch, query and report events are processed by workqueue
> So context is sleepable, not atomic.
> While this logic, it acquires RTNL.
> 
> 7. Add new mc_lock.
> The purpose of this lock is to protect per-interface mld data.
> Per-interface mld data is usually used by query/report event handler.
> So, query/report event workers need only this lock instead of RTNL.
> Therefore, it could reduce bottleneck.
> 
> Changelog:
> v2 -> v3:
> 1. Do not use msecs_to_jiffies().
> (by Cong Wang)
> 2. Do not add unnecessary rtnl_lock() and rtnl_unlock().
> (by Cong Wang)
> 3. Fix sparse warnings because of rcu annotation.
> (by kernel test robot)
>     - Remove some rcu_assign_pointer(), which was used for non-rcu pointer.
>     - Add union for rcu pointer.
>     - Use rcu API in mld_clear_zeros().
>     - Remove remained rcu_read_unlock().
>     - Use rcu API for tomb resources.
> 4. withdraw prevopus 2nd and 3rd patch.
>     - "separate two flags from ifmcaddr6->mca_flags"
>     - "add a new delayed_work, mc_delrec_work"
> 5. Add 6th and 7th patch.
> 
> v1 -> v2:
> 1. Withdraw unnecessary refactoring patches.
> (by Cong Wang, Eric Dumazet, David Ahern)
>      a) convert from array to list.
>      b) function rename.
> 2. Separate big one patch into small several patches.
> 3. Do not rename 'ifmcaddr6->mca_lock'.
> In the v1 patch, this variable was changed to 'ifmcaddr6->mca_work_lock'.
> But this is actually not needed.
> 4. Do not use atomic_t for 'ifmcaddr6->mca_sfcount' and
> 'ipv6_mc_socklist'->sf_count'.
> 5. Do not add mld_check_leave_group() function.
> 6. Do not add ip6_mc_del_src_bulk() function.
> 7. Do not add ip6_mc_add_src_bulk() function.
> 8. Do not use rcu_read_lock() in the qeth_l3_add_mcast_rtnl().
> (by Julian Wiedmann)
> 
> Taehee Yoo (7):
>    mld: convert from timer to delayed work
>    mld: get rid of inet6_dev->mc_lock
>    mld: convert ipv6_mc_socklist->sflist to RCU
>    mld: convert ip6_sf_list to RCU
>    mld: convert ifmcaddr6 to RCU
>    mld: add new workqueues for process mld events
>    mld: add mc_lock for protecting per-interface mld data
> 
>   drivers/s390/net/qeth_l3_main.c |    6 +-
>   include/net/if_inet6.h          |   37 +-
>   include/net/mld.h               |    3 +
>   net/batman-adv/multicast.c      |    6 +-
>   net/ipv6/addrconf.c             |    9 +-
>   net/ipv6/addrconf_core.c        |    2 +-
>   net/ipv6/af_inet6.c             |    2 +-
>   net/ipv6/icmp.c                 |    4 +-
>   net/ipv6/mcast.c                | 1080 ++++++++++++++++++-------------
>   9 files changed, 678 insertions(+), 471 deletions(-)
> 

Powered by blists - more mailing lists