netdev - Re: [PATCH] page_pool: use irqsave/irqrestore to protect ring access.

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Tue, 10 Mar 2020 23:34:16 +0000
From:   Saeed Mahameed <saeedm@...lanox.com>
To:     "jonathan.lemon@...il.com" <jonathan.lemon@...il.com>
CC:     "kernel-team@...com" <kernel-team@...com>,
        "davem@...emloft.net" <davem@...emloft.net>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "brouer@...hat.com" <brouer@...hat.com>,
        "ilias.apalodimas@...aro.org" <ilias.apalodimas@...aro.org>
Subject: Re: [PATCH] page_pool: use irqsave/irqrestore to protect ring access.

On Tue, 2020-03-10 at 10:07 -0700, Jonathan Lemon wrote:
> 
> On 9 Mar 2020, at 19:30, Saeed Mahameed wrote:
> 
> > On Mon, 2020-03-09 at 17:55 -0700, David Miller wrote:
> > > From: Jonathan Lemon <jonathan.lemon@...il.com>
> > > Date: Mon, 9 Mar 2020 12:49:29 -0700
> > > 
> > > > netpoll may be called from IRQ context, which may access the
> > > > page pool ring.  The current _bh variants do not provide
> > > > sufficient
> > > > protection, so use irqsave/restore instead.
> > > > 
> > > > Error observed on a modified mlx4 driver, but the code path
> > > > exists
> > > > for any driver which calls page_pool_recycle from napi poll.
> > > > 
> > > > WARNING: CPU: 34 PID: 550248 at /ro/source/kernel/softirq.c:161
> > > __local_bh_enable_ip+0x35/0x50
> > >  ...
> > > > Signed-off-by: Jonathan Lemon <jonathan.lemon@...il.com>
> > > 
> > > The netpoll stuff always makes the locking more complicated than
> > > it
> > > needs
> > > to be.  I wonder if there is another way around this issue?
> > > 
> > > Because IRQ save/restore is a high cost to pay in this critical
> > > path.
> > 
> > a printk inside irq context lead to this, so maybe it can be
> > avoided ..
> 
> This was caused by a printk in hpet_rtc_timer_reinit() complaining
> about
> RTC interrupts being lost.  I'm not sure it's practical trying to
> locate
> all the printk cases like this.
> 
> 
> > or instead of checking in_serving_softirq()  change page_pool to
> > check in_interrupt() which is more powerful, to avoid ptr_ring
> > locking
> > and the complication with netpoll altogether.
> 
> That's another approach:
> 
>     ret = 1;
>     if (!in_irq()) {
>         if (in_serving_softirq())
>             ret = ptr_ring_produce(....
>         else
>             ret = ptr_ring_produce_bh(....
>     }
> 
> which would return failure and release the page from the page pool.
> This doesn't address the allocation or the bulk release path.
> 
> 
> > I wonder why Jesper picked in_serving_softirq() in first place, was
> > there a specific reason ? or he just wanted it to be as less strict
> > as
> > possible ?
> 
> From the code, it looks like he was optimizing to avoid the _bh
> variant
> if possible.

I think that if you use another variant this it intorduce a possiblerace bestrewn sofitq and hardirq ..