netdev - Re: [RFC next-next v2 1/5] net: marvell: prestera: Add driver for Prestera family ASIC devices

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]

Message-ID: <20200512071552.GA17235@plvision.eu>
Date:   Tue, 12 May 2020 10:15:52 +0300
From:   Vadym Kochan <vadym.kochan@...ision.eu>
To:     Jiri Pirko <jiri@...nulli.us>
Cc:     netdev@...r.kernel.org, "David S. Miller" <davem@...emloft.net>,
        Oleksandr Mazur <oleksandr.mazur@...ision.eu>,
        Serhiy Boiko <serhiy.boiko@...ision.eu>,
        Serhiy Pshyk <serhiy.pshyk@...ision.eu>,
        Volodymyr Mytnyk <volodymyr.mytnyk@...ision.eu>,
        Taras Chornyi <taras.chornyi@...ision.eu>,
        Andrii Savka <andrii.savka@...ision.eu>,
        Jiri Pirko <jiri@...lanox.com>,
        Ido Schimmel <idosch@...lanox.com>,
        Andrew Lunn <andrew@...n.ch>,
        Chris Packham <Chris.Packham@...iedtelesis.co.nz>
Subject: Re: [RFC next-next v2 1/5] net: marvell: prestera: Add driver for
 Prestera family ASIC devices

On Tue, May 12, 2020 at 07:55:36AM +0200, Jiri Pirko wrote:
> Mon, May 11, 2020 at 09:24:22PM CEST, vadym.kochan@...ision.eu wrote:
> >On Mon, May 11, 2020 at 02:57:23PM +0200, Jiri Pirko wrote:
> >> [...]
> >> 
> >> >diff --git a/drivers/net/ethernet/marvell/prestera/prestera_dsa.c b/drivers/net/ethernet/marvell/prestera/prestera_dsa.c
[...]
> >> >+netdev_tx_t prestera_sdma_xmit(struct prestera_sdma *sdma, struct sk_buff *skb)
> >> >+{
> >> >+	struct device *dma_dev = sdma->sw->dev->dev;
> >> >+	struct prestera_tx_ring *tx_ring;
> >> >+	struct net_device *dev = skb->dev;
> >> >+	struct prestera_sdma_buf *buf;
> >> >+	int err;
> >> >+
> >> >+	tx_ring = &sdma->tx_ring;
> >> >+
> >> >+	buf = &tx_ring->bufs[tx_ring->next_tx];
> >> >+	if (buf->is_used) {
> >> >+		schedule_work(&sdma->tx_work);
> >> >+		goto drop_skb;
> >> >+	}
> >> 
> >> What is preventing 2 CPUs to get here and work with the same buf?
> >
> >I assume you mean serialization between the recycling work and xmit
> >context ? Actually they are just updating 'is_used' field which
> 
> No.
> 
> >allows to use or free, what I can see is that may be I need to use
> >something like READ_ONCE/WRITE_ONCE, but the rest looks safe for me:
> >
> >1) recycler updates is_used=false only after fully freeing the buffer,
> >and only if it was set to true.
> >
> >2) xmit context gets next buffer to use only if it is freed
> >(is_used=false), and sets it to true after buffer is ready to be sent.
> >
> >So, yes these contexts both update this field but in strict sequence.
> >
> >If you mean of protecting of xmit on several CPUS so, the xmit should be
> >serialized on kernel, and the driver uses one queue which (as I
> >underand) is bound to particular CPU.
> 
> How is it serialized? You get here (to prestera_sdma_xmit()) on 2 CPUs
> with the same sdma pointer and 2 skbs.
> 

My understanding is:

dev_hard_start_xmit is the entry function which is called by the
networking layer to send skb via device (qos scheduler, pktgen, xfrm,
core - dev_direct_xmit(), etc).

All they acquire the HARD_TX_LOCK which locks particular tx queue. And
since the driver uses one tx queue there should be no concurrent access
inside ndo_start_xmit, right ?

[...]