Re: [PATCH v4] irqchip/ls-extirq: fix invalid wait context by avoiding to use regmap

From: Marc Zyngier
Date: Thu Jul 28 2022 - 13:32:36 EST


Sean

Please don't mix the threads. This makes it impossible to follow.

On Thu, 28 Jul 2022 16:43:40 +0100,
Sean Anderson <sean.anderson@xxxxxxxx> wrote:
>
>
>
> On 7/28/22 11:28 AM, Vladimir Oltean wrote:
> > On Thu, Jul 28, 2022 at 11:25:23AM -0400, Sean Anderson wrote:
> >> Could we just use use_raw_spinlock in the regmap config?
> >
> > That was v2, essentially:
> > https://lore.kernel.org/lkml/874jz6dcp6.wl-maz@xxxxxxxxxx/
> >
>
> On 7/24/22 6:31 AM, Marc Zyngier wrote:
> > On Fri, 22 Jul 2022 21:40:19 +0100,
> > Vladimir Oltean <vladimir.oltean@xxxxxxx> wrote:
> >>
> >> The irqchip->irq_set_type method is called by __irq_set_trigger() under
> >> the desc->lock raw spinlock.
> >>
> >> The ls-extirq implementation, ls_extirq_irq_set_type(), uses an MMIO
> >> regmap created by of_syscon_register(), which uses plain spinlocks
> >> (the kind that are sleepable on RT).
> >>
> >> Therefore, this is an invalid locking scheme for which we get a kernel
> >> splat stating just that ("[ BUG: Invalid wait context ]"), because the
> >> context in which the plain spinlock may sleep is atomic due to the raw
> >> spinlock. We need to go raw spinlocks all the way.
> >>
> >> Make this driver create its own MMIO regmap, with use_raw_spinlock=true,
> >> and stop relying on syscon to provide it. Since the regmap we got from
> >> syscon belonged to the parent and this one belongs to us, the offset to
> >> the INTPCR register is now 0, because of the address translation that
> >> takes place through the device tree.
> >>
> >> Another complication that we need to deal with is the fact that we need
> >> to parse the 'little-endian'/'big-endian' specifiers that exist in
> >> device trees for the parent ourselves now, since we no longer involve
> >> syscon.
> >>
> >> And yet one final complication, due to the fact that this driver uses
> >> IRQCHIP_DECLARE rather than traditional platform devices with probe and
> >> remove methods, is that we cannot use devres, so we need to implement a
> >> full-blown cleanup procedure on the error path.
> >>
> >> This patch depends on commit 67021f25d952 ("regmap: teach regmap to use
> >> raw spinlocks if requested in the config").
> >
> > This information doesn't belong to the commit message (and please read
> > the documentation about the use of "This patch").
> >
> >>
> >> Fixes: 0dcd9f872769 ("irqchip: Add support for Layerscape external interrupt lines")
> >> Signed-off-by: Vladimir Oltean <vladimir.oltean@xxxxxxx>
> >> ---
> >> v1->v2: create a separate regmap for the ls-extirq driver rather than
> >> relying on the one provided by syscon or modifying that.
> >>
> >> For reference, v1 is at:
> >> https://lore.kernel.org/lkml/20210825205041.927788-3-vladimir.oltean@xxxxxxx/
> >>
> >> For extra reviewer convenience, the ls-extirq appears in the following
> >> SoC device trees:
> >> https://elixir.bootlin.com/linux/v5.18.13/source/arch/arm64/boot/dts/freescale/fsl-ls208xa.dtsi#L289
> >> https://elixir.bootlin.com/linux/v5.18.13/source/arch/arm64/boot/dts/freescale/fsl-ls1088a.dtsi#L249
> >> https://elixir.bootlin.com/linux/v5.18.13/source/arch/arm64/boot/dts/freescale/fsl-ls1043a.dtsi#L319
> >> https://elixir.bootlin.com/linux/v5.18.13/source/arch/arm64/boot/dts/freescale/fsl-ls1046a.dtsi#L325
> >> https://elixir.bootlin.com/linux/v5.18.13/source/arch/arm64/boot/dts/freescale/fsl-lx2160a.dtsi#L682
> >> https://elixir.bootlin.com/linux/v5.18.13/source/arch/arm/boot/dts/ls1021a.dtsi#L182
> >>
> >> Patch tested on LX2160A and LS1021A.
> >>
> >> drivers/irqchip/irq-ls-extirq.c | 81 +++++++++++++++++++++++----------
> >> 1 file changed, 58 insertions(+), 23 deletions(-)
> >>
> >> diff --git a/drivers/irqchip/irq-ls-extirq.c b/drivers/irqchip/irq-ls-extirq.c
> >> index 853b3972dbe7..94a22642b3f2 100644
> >> --- a/drivers/irqchip/irq-ls-extirq.c
> >> +++ b/drivers/irqchip/irq-ls-extirq.c
> >> @@ -6,7 +6,7 @@
> >> #include <linux/irqchip.h>
> >> #include <linux/irqdomain.h>
> >> #include <linux/of.h>
> >> -#include <linux/mfd/syscon.h>
> >> +#include <linux/of_address.h>
> >> #include <linux/regmap.h>
> >> #include <linux/slab.h>
> >>
> >> @@ -16,8 +16,7 @@
> >> #define LS1021A_SCFGREVCR 0x200
> >>
> >> struct ls_extirq_data {
> >> - struct regmap *syscon;
> >> - u32 intpcr;
> >> + struct regmap *regmap;
> >> bool is_ls1021a_or_ls1043a;
> >> u32 nirq;
> >> struct irq_fwspec map[MAXIRQ];
> >> @@ -51,7 +50,10 @@ ls_extirq_set_type(struct irq_data *data, unsigned int type)
> >> default:
> >> return -EINVAL;
> >> }
> >> - regmap_update_bits(priv->syscon, priv->intpcr, mask, value);
> >> + /* INTPCR is the only register of our regmap,
> >> + * therefore its offset is 0
> >> + */
> >
> > For multi-line comments, please use the normal kernel comment style,
> > not the one that is mandated for net.
> >
> >> + regmap_update_bits(priv->regmap, 0, mask, value);
> >>
> >> return irq_chip_set_type_parent(data, type);
> >> }
> >> @@ -143,48 +145,81 @@ ls_extirq_parse_map(struct ls_extirq_data *priv, struct device_node *node)
> >> static int __init
> >> ls_extirq_of_init(struct device_node *node, struct device_node *parent)
> >> {
> >> -
> >> + struct regmap_config extirq_regmap_config = {
> >> + .name = "intpcr",
> >> + .reg_bits = 32,
> >> + .val_bits = 32,
> >> + .reg_stride = 4,
> >> + .use_raw_spinlock = true,
> >> + };
> >> struct irq_domain *domain, *parent_domain;
> >> struct ls_extirq_data *priv;
> >> + void __iomem *base;
> >> int ret;
> >>
> >> parent_domain = irq_find_host(parent);
> >> if (!parent_domain) {
> >> pr_err("Cannot find parent domain\n");
> >> - return -ENODEV;
> >> + ret = -ENODEV;
> >> + goto err_irq_find_host;
> >> }
> >>
> >> priv = kzalloc(sizeof(*priv), GFP_KERNEL);
> >> - if (!priv)
> >> - return -ENOMEM;
> >> -
> >> - priv->syscon = syscon_node_to_regmap(node->parent);
> >> - if (IS_ERR(priv->syscon)) {
> >> - ret = PTR_ERR(priv->syscon);
> >> - pr_err("Failed to lookup parent regmap\n");
> >> - goto out;
> >> + if (!priv) {
> >> + ret = -ENOMEM;
> >> + goto err_alloc_priv;
> >> + }
> >> +
> >> + /* All extirq OF nodes are under a scfg/syscon node with
> >> + * the 'ranges' property
> >> + */
> >> + base = of_iomap(node, 0);
> >> + if (!base) {
> >> + pr_err("Cannot ioremap OF node %pOF\n", node);
> >> + ret = -ENOMEM;
> >> + goto err_iomap;
> >> }
> >> - ret = of_property_read_u32(node, "reg", &priv->intpcr);
> >> - if (ret) {
> >> - pr_err("Missing INTPCR offset value\n");
> >> - goto out;
> >> +
> >> + /* Parse the parent device's DT node for an endianness specification */
> >> + if (of_property_read_bool(parent, "big-endian"))
> >> + extirq_regmap_config.val_format_endian = REGMAP_ENDIAN_BIG;
> >> + else if (of_property_read_bool(parent, "little-endian"))
> >> + extirq_regmap_config.val_format_endian = REGMAP_ENDIAN_LITTLE;
> >> + else if (of_property_read_bool(parent, "native-endian"))
> >> + extirq_regmap_config.val_format_endian = REGMAP_ENDIAN_NATIVE;
> >
> > All of this should be rewritten to use of_device_is_big_endian(), and
> > reduce the whole thing to two cases (I don't think native endian makes
> > much sense anyway). I also wonder what the result is if none of these
> > properties is present...
>
> I think regmap_get_val_endian would be better here.

Why?

>
> >> +
> >> + priv->regmap = regmap_init_mmio(NULL, base, &extirq_regmap_config);
>
> It could also be done automatically if we pass the syscon dev instead of
> NULL. The only downside is that some regmap error messages will use the
> syscon device
>
> >> + if (IS_ERR(priv->regmap)) {
> >> + pr_err("Cannot create MMIO regmap: %pe\n", priv->regmap);
> >> + ret = PTR_ERR(priv->regmap);
> >> + goto err_regmap_init;
> >
> > Finally, what is the actual benefit of using a regmap here? It seems
> > like a very roundabout way of performing a RMW on a register whilst
> > holding a lock... Passing NULL for a device to regmap_init_mmio() also
> > seems to be an extremely rare idiom (only 5 cases in the tree), and
> > this doesn't seem completely right to me.
>
> The benefit is that you don't have to write (yet another) set of
> endian-converting read/write functions. The above (non-NULL) usage of
> regmap_init would also address your criticism here.

I really don't see the point of adding yet another abstraction layer
for something can be efficiently be managed locally.

regmap is good at managing things that are repeatedly updated, and
keep the most up to date version. This isn't the case here. This
register is read/written *once* per interrupt being configured. There
is no performance requirement.

So no, I don't think we need this at all.

M.

--
Without deviation from the norm, progress is not possible.