Message-ID: <20250211114307.GB5888@pendragon.ideasonboard.com>
Date: Tue, 11 Feb 2025 13:43:07 +0200
From: Laurent Pinchart <laurent.pinchart@...asonboard.com>
To: "zhangzekun (A)" <zhangzekun11@...wei.com>
Cc: Oleksij Rempel <o.rempel@...gutronix.de>, robh@...nel.org,
	saravanak@...gle.com, justin.chen@...adcom.com,
	florian.fainelli@...adcom.com, andrew+netdev@...n.ch,
	kuba@...nel.org, kory.maincent@...tlin.com,
	jacopo+renesas@...ndi.org, kieran.bingham+renesas@...asonboard.com,
	maddy@...ux.ibm.com, mpe@...erman.id.au, npiggin@...il.com,
	olteanv@...il.com, davem@...emloft.net, taras.chornyi@...ision.eu,
	edumazet@...gle.com, pabeni@...hat.com, sudeep.holla@....com,
	cristian.marussi@....com, arm-scmi@...r.kernel.org,
	linuxppc-dev@...ts.ozlabs.org, linux-media@...r.kernel.org,
	netdev@...r.kernel.org, devicetree@...r.kernel.org,
	chenjun102@...wei.com
Subject: Re: [PATCH 1/9] of: Add warpper function
 of_find_node_by_name_balanced()

On Tue, Feb 11, 2025 at 07:26:18PM +0800, zhangzekun (A) wrote:
> 在 2025/2/10 18:03, Laurent Pinchart 写道:
> > On Mon, Feb 10, 2025 at 02:47:28PM +0800, zhangzekun (A) wrote:
> >>> I think we all agree that of_find_node_by_name() is miused, and that it
> >>> shows the API isn't optimal. What we have different opinions on is how
> >>> to make the API less error-prone. I think adding a new
> >>> of_find_node_by_name_balanced() function works around the issue and
> >>> doesn't improve the situation much, I would argue it makes things even
> >>> more confusing.
> >>>
> >>> We have only 20 calls to of_find_node_by_name() with a non-NULL first
> >>> argument in v6.14-rc1:
> >>>
> >>> arch/powerpc/platforms/chrp/pci.c:      rtas = of_find_node_by_name (root, "rtas");
> >>>
> >>> The 'root' variable here is the result of a call to
> >>> 'of_find_node_by_path("/")', so I think we could pass a null pointer
> >>> instead to simplify things.
> >>>
> >>> arch/powerpc/platforms/powermac/pic.c:          slave = of_find_node_by_name(master, "mac-io");
> >>>
> >>> Here I believe of_find_node_by_name() is called to find a *child* node
> >>> of 'master'. of_find_node_by_name() is the wrong function for that.
> >>>
> >>> arch/sparc/kernel/leon_kernel.c:        np = of_find_node_by_name(rootnp, "GAISLER_IRQMP");
> >>> arch/sparc/kernel/leon_kernel.c:                np = of_find_node_by_name(rootnp, "01_00d");
> >>> arch/sparc/kernel/leon_kernel.c:        np = of_find_node_by_name(nnp, "GAISLER_GPTIMER");
> >>> arch/sparc/kernel/leon_kernel.c:                np = of_find_node_by_name(nnp, "01_011");
> >>>
> >>> Here too the code seems to be looking for child nodes only (but I
> >>> couldn't find a DT example or binding in-tree, so I'm not entirely
> >>> sure).
> >>>
> >>> drivers/clk/ti/clk.c:   return of_find_node_by_name(from, tmp);
> >>>
> >>> Usage here seems correct, the reference-count decrement is intended.
> >>>
> >>> drivers/media/i2c/max9286.c:    i2c_mux = of_find_node_by_name(dev->of_node, "i2c-mux");
> >>> drivers/media/platform/qcom/venus/core.c:       enp = of_find_node_by_name(dev->of_node, node_name);
> >>> drivers/net/dsa/bcm_sf2.c:      ports = of_find_node_by_name(dn, "ports");
> >>> drivers/net/dsa/hirschmann/hellcreek_ptp.c:     leds = of_find_node_by_name(hellcreek->dev->of_node, "leds");
> >>> drivers/net/ethernet/broadcom/asp2/bcmasp.c:    ports_node = of_find_node_by_name(dev->of_node, "ethernet-ports");
> >>> drivers/net/ethernet/marvell/prestera/prestera_main.c:  ports = of_find_node_by_name(sw->np, "ports");
> >>> drivers/net/pse-pd/tps23881.c:  channels_node = of_find_node_by_name(priv->np, "channels");
> >>> drivers/regulator/scmi-regulator.c:     np = of_find_node_by_name(handle->dev->of_node, "regulators");
> >>> drivers/regulator/tps6594-regulator.c:          np = of_find_node_by_name(tps->dev->of_node, multi_regs[multi].supply_name);
> >>>
> >>> Incorrect usage, as far as I understand all those drivers are looking
> >>> for child nodes only.
> >>>
> >>> drivers/of/unittest.c:          found = of_find_node_by_name(nd->overlay, "test-unittest16");
> >>> drivers/of/unittest.c:          found = of_find_node_by_name(nd->overlay, "test-unittest17");
> >>> drivers/of/unittest.c:          found = of_find_node_by_name(nd->overlay, "test-unittest18");
> >>> drivers/of/unittest.c:          found = of_find_node_by_name(nd->overlay, "test-unittest19");
> >>>
> >>> Here too I think only child nodes are meant to be considered.
> >>>
> >>> of_find_node_by_name() is very much misused as most callers want to find
> >>> child nodes, while of_find_node_by_name() will walk the whole DT from a
> >>> given starting point.
> >>>
> >>> I think the right fix here is to
> >>>
> >>> - Replace of_find_node_by_name(root, ...) with
> >>>     of_find_node_by_name(NULL, ...) in arch/powerpc/platforms/chrp/pci.c
> >>>     (if my understanding of the code is correct).
> >>
> >> For arch/powerpc/platforms/chrp/pci.c, noticing that there is a comment
> >> in setup_peg2():
> >>    /* keep the reference to the root node */
> >>
> >> It might can not be convert to of_find_node_by_name(NULL, ...), and the
> >> origin use of of_find_node_by_name() put the ref count which want to be
> >> kept.
> > 
> > But the reference is dropped by of_find_node_by_name(). Unless I'm
> > missing something, dropping the lien
> > 
> > 	struct device_node *root = of_find_node_by_path("/");
> > 
> > and changing
> > 
> > 	rtas = of_find_node_by_name (root, "rtas");
> > 
> > to
> > 
> > 	rtas = of_find_node_by_name (NULL, "rtas");
> > 
> > will not change the behaviour of the code.
> 
> Hi, Laurent,
> 
> I think that the original code try to keep the refcount get by 
> of_find_node_by_path(), but leak it accidently by 
> of_find_node_by_name(). I am not sure that what driver really wants to 
> do and if it has a bug here.

Looking at the git history, I don't think the code needs or tries to
keep a reference to the root node.

-- 
Regards,

Laurent Pinchart

Powered by blists - more mailing lists