Re: [PATCH] irqchip/irq-mtk-sysirq: drop unnecessary spinlock

[Marc Zyngier]

On 2020-05-27 17:17, Bartosz Golaszewski wrote:
> From: Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>
>
> This driver takes a regular spinlock when a raw spinlock is already
> taken which results in the following lockdep splat:
>
> =============================
> [ BUG: Invalid wait context ]
> 5.7.0-rc6-02446-gb9827c0a9fe7-dirty #1 Not tainted
> -----------------------------
> swapper/0/0 is trying to lock:
> ffffff800303b798 (&chip_data->lock){....}-{3:3}, at:
> mtk_sysirq_set_type+0x48/0xc0
> other info that might help us debug this:
> context-{5:5}
> 2 locks held by swapper/0/0:
>  #0: ffffff800302ee68 (&desc->request_mutex){....}-{4:4}, at:
> __setup_irq+0xc4/0x8a0
>  #1: ffffff800302ecf0 (&irq_desc_lock_class){....}-{2:2}, at:
> __setup_irq+0xe4/0x8a0
> stack backtrace:
> CPU: 0 PID: 0 Comm: swapper/0 Not tainted 
> 5.7.0-rc6-02446-gb9827c0a9fe7-dirty #1
> Hardware name: Pumpkin MT8516 (DT)
> Call trace:
>  dump_backtrace+0x0/0x180
>  show_stack+0x14/0x20
>  dump_stack+0xd0/0x118
>  __lock_acquire+0x8c8/0x2270
>  lock_acquire+0xf8/0x470
>  _raw_spin_lock_irqsave+0x50/0x78
>  mtk_sysirq_set_type+0x48/0xc0
>  __irq_set_trigger+0x58/0x170
>  __setup_irq+0x420/0x8a0
>  request_threaded_irq+0xd8/0x190
>  timer_of_init+0x1e8/0x2c4
>  mtk_gpt_init+0x5c/0x1dc
>  timer_probe+0x74/0xf4
>  time_init+0x14/0x44
>  start_kernel+0x394/0x4f0
>
> We don't need the spinlock here - the irq_set_type() callback is always
> called with the irq_desc->lock taken. This removes the spinlock 
> entirely.

It looks really great. Not.

> Signed-off-by: Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>
> ---
>  drivers/irqchip/irq-mtk-sysirq.c | 6 ------
>  1 file changed, 6 deletions(-)
>
> diff --git a/drivers/irqchip/irq-mtk-sysirq.c 
> b/drivers/irqchip/irq-mtk-sysirq.c
> index 73eae5966a40..da2fc4809222 100644
> --- a/drivers/irqchip/irq-mtk-sysirq.c
> +++ b/drivers/irqchip/irq-mtk-sysirq.c
> @@ -12,10 +12,8 @@
>  #include <linux/of_address.h>
>  #include <linux/io.h>
>  #include <linux/slab.h>
> -#include <linux/spinlock.h>
>
>  struct mtk_sysirq_chip_data {
> -	spinlock_t lock;
>  	u32 nr_intpol_bases;
>  	void __iomem **intpol_bases;
>  	u32 *intpol_words;
> @@ -30,14 +28,12 @@ static int mtk_sysirq_set_type(struct irq_data
> *data, unsigned int type)
>  	u8 intpol_idx = chip_data->intpol_idx[hwirq];
>  	void __iomem *base;
>  	u32 offset, reg_index, value;
> -	unsigned long flags;
>  	int ret;
>
>  	base = chip_data->intpol_bases[intpol_idx];
>  	reg_index = chip_data->which_word[hwirq];
>  	offset = hwirq & 0x1f;
>
> -	spin_lock_irqsave(&chip_data->lock, flags);
>  	value = readl_relaxed(base + reg_index * 4);
>  	if (type == IRQ_TYPE_LEVEL_LOW || type == IRQ_TYPE_EDGE_FALLING) {
>  		if (type == IRQ_TYPE_LEVEL_LOW)
> @@ -53,7 +49,6 @@ static int mtk_sysirq_set_type(struct irq_data
> *data, unsigned int type)
>
>  	data = data->parent_data;
>  	ret = data->chip->irq_set_type(data, type);
> -	spin_unlock_irqrestore(&chip_data->lock, flags);
>  	return ret;
>  }
>
> @@ -212,7 +207,6 @@ static int __init mtk_sysirq_of_init(struct
> device_node *node,
>  		ret = -ENOMEM;
>  		goto out_free_which_word;
>  	}
> -	spin_lock_init(&chip_data->lock);
>
>  	return 0;

Sight... Do you realize that these two locks do not protect the same
thing at all? One protects the interrupt data, and the other protects
the MMIO register which is shared between multiple interrupts, and
on which the driver performs a RMW.

Thanks to the removal of this spinlock, two irq_set_type() can execute
in parallel and silently corrupt the register. Not exactly an 
improvement.

        M.
--
Jazz is not dead. It just smells funny...