Message-ID: <YZynSa6s8kBKtSYB@unreal>
Date:   Tue, 23 Nov 2021 10:33:13 +0200
From:   Leon Romanovsky <leon@...nel.org>
To:     Jakub Kicinski <kuba@...nel.org>
Cc:     "David S . Miller" <davem@...emloft.net>,
        Alexandre Belloni <alexandre.belloni@...tlin.com>,
        Andrew Lunn <andrew@...n.ch>, Aya Levin <ayal@...lanox.com>,
        Claudiu Manoil <claudiu.manoil@....com>, drivers@...sando.io,
        Florian Fainelli <f.fainelli@...il.com>,
        Ido Schimmel <idosch@...dia.com>,
        intel-wired-lan@...ts.osuosl.org,
        Ioana Ciornei <ioana.ciornei@....com>,
        Jesse Brandeburg <jesse.brandeburg@...el.com>,
        Jiri Pirko <jiri@...dia.com>, linux-kernel@...r.kernel.org,
        linux-rdma@...r.kernel.org,
        Michael Chan <michael.chan@...adcom.com>,
        netdev@...r.kernel.org, oss-drivers@...igine.com,
        Saeed Mahameed <saeedm@...dia.com>,
        Shannon Nelson <snelson@...sando.io>,
        Simon Horman <simon.horman@...igine.com>,
        Taras Chornyi <tchornyi@...vell.com>,
        Tariq Toukan <tariqt@...dia.com>,
        Tony Nguyen <anthony.l.nguyen@...el.com>,
        UNGLinuxDriver@...rochip.com,
        Vivien Didelot <vivien.didelot@...il.com>,
        Vladimir Oltean <vladimir.oltean@....com>
Subject: Re: [PATCH net-next 5/6] devlink: Reshuffle resource registration
 logic

On Mon, Nov 22, 2021 at 06:27:28PM -0800, Jakub Kicinski wrote:
> On Sun, 21 Nov 2021 10:45:12 +0200 Leon Romanovsky wrote:
> > On Fri, Nov 19, 2021 at 08:10:17AM -0800, Jakub Kicinski wrote:
> > > On Fri, 19 Nov 2021 17:38:53 +0200 Leon Romanovsky wrote:  
> > > > My approach works, exactly like it works in other subsystems.
> > > > https://lore.kernel.org/netdev/cover.1636390483.git.leonro@nvidia.com/  
> > > 
> > > What "other subsystems"? I'm aware of the RFC version of these patches.  
> > 
> > Approach to have fine-grained locking scheme, instead of having one big lock.
> > This was done in MM for mmap_sem, we did it for RDMA too.
> 
> You're breaking things up to avoid lock ordering issues. The user can
> still only run a single write command at a time.
> 
> > > Breaking up the locks to to protect sub-objects only is fine for
> > > protecting internal lists but now you can't guarantee that the object
> > > exists when driver is called.  
> > 
> > I can only guess about which objects you are talking.
> 
> It obviously refers to the port splitting I mentioned below.
> 
> > If you are talking about various devlink sub-objects (ports, traps,
> > e.t.c), they created by the drivers and as such should be managed by them.
> > Also they are connected to devlink which is guaranteed to exist. At the end,
> > they called to devlink_XXX->devlink pointer without any existence check.
> > 
> > If you are talking about devlink instance itself, we guarantee that it
> > exists between devlink_alloc() and devlink_free(). It seems to me pretty
> > reasonable request from drivers do not access devlink before devlink_alloc()
> > or after devlink_free(),
> > 
> > > I'm sure you'll utter your unprovable "in real drivers.." but the fact
> > > is my approach does not suffer from any such issues. Or depends on
> > > drivers registering devlink last.  
> > 
> > Registration of devlink doesn't do anything except opening it to the world.
> > The lifetime is controlled with alloc and free. My beloved sentence "in
> > real drivers ..." belongs to use of devlink_put and devlink_locks outside
> > of devlink.c and nothing more.
> 
> As soon as there is a inter-dependency between two subsystems "must 
> be last" breaks down.

"Must be last" is better to be changed "when the device is ready".

-----------------------------------------------------------------
> The real question is whether you now require devlink_register()
> to go last in general? 

No, it is not requirement, but my suggestion. You need to be aware that
after call to devlink_register(), the device will be fully open for devlink
netlink access. So it is strongly advised to put devlink_register to be the
last command in PCI initialization sequence.

https://lore.kernel.org/netdev/YXhVd16heaHCegL1@unreal/
--------------------------------------------------------------------

> 
> > > I can start passing a pointer to a devlink_port to split/unsplit
> > > functions, which is a great improvement to the devlink driver API.  
> > 
> > You can do it with my approach too. We incremented reference counter
> > of devlink instance when devlink_nl_cmd_port_split_doit() was called,
> > and we can safely take devlink->port_list_lock lock before returning
> > from pre_doit.
> 
> Wait, I thought you'd hold devlink->lock around split/unsplit.

I'm holding.

    519 static int devlink_nl_pre_doit(const struct genl_ops *ops,
    520                                struct sk_buff *skb, struct genl_info *info)
    521 {
    ...
    529
    530         mutex_lock(&devlink->lock);


> 
> Please look at the port splitting case, mlx5 doesn't implement it
> but it's an important feature.

I'll, but please don't forget that it was RFC, just to present that
devlink can be changed internally without exposing internals.

> 
> Either way, IDK how ref count on devlink helps with lifetime of a
> subobject. You must assume the sub-objects can only be created outside
> of the time devlink instance is visible or under devlink->lock?

The devlink lifetime is:
stages:        I                   II                   III   
 devlink_alloc -> devlink_register -> devlink_unregister -> devlink_free.

All sub-objects should be created between devlink_alloc and devlink_free.
It will ensure that ->devlink pointer is always valid.

Stage I:
 * There is no need to hold any devlink locks or increase reference counter.
   If driver doesn't do anything crazy during its init, nothing in devlink
   land will run in parallel. 
Stage II:
 * There is a need to hold devlink->lock and/or play with reference counter
   and/or use fine-grained locks. Users can issue "devlink ..." commands.
Stage III:

Thanks

Powered by blists - more mailing lists