Message-ID: <CAC5umyjdM1+4nPg_6UaCjcpikESamdA_ZpmP4Xfjx7_-1f=+0A@mail.gmail.com>
Date:   Sat, 28 Sep 2019 01:47:21 +0900
From:   Akinobu Mita <akinobu.mita@...il.com>
To:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>
Cc:     linux-leds@...r.kernel.org, LKML <linux-kernel@...r.kernel.org>,
        "Rafael J. Wysocki" <rafael@...nel.org>,
        Jacek Anaszewski <jacek.anaszewski@...il.com>,
        Pavel Machek <pavel@....cz>, Dan Murphy <dmurphy@...com>
Subject: Re: [PATCH v2 1/1] leds: remove PAGE_SIZE limit of /sys/class/leds/<led>/trigger

2019年9月27日(金) 15:39 Greg Kroah-Hartman <gregkh@...uxfoundation.org>:
>
> On Sat, Sep 14, 2019 at 12:03:24AM +0900, Akinobu Mita wrote:
> > Reading /sys/class/leds/<led>/trigger returns all available LED triggers.
> > However, the size of this file is limited to PAGE_SIZE because of the
> > limitation for sysfs attribute.
> >
> > Enabling LED CPU trigger on systems with thousands of CPUs easily hits
> > PAGE_SIZE limit, and makes it impossible to see all available LED triggers
> > and which trigger is currently activated.
> >
> > We work around it here by converting /sys/class/leds/<led>/trigger to
> > binary attribute, which is not limited by length. This is _not_ good
> > design, do not copy it.
> >
> > Cc: Greg Kroah-Hartman <gregkh@...uxfoundation.org>
> > Cc: "Rafael J. Wysocki" <rafael@...nel.org>
> > Cc: Jacek Anaszewski <jacek.anaszewski@...il.com>
> > Cc: Pavel Machek <pavel@....cz>
> > Cc: Dan Murphy <dmurphy@...com>
> > Acked-by: Pavel Machek <pavel@....cz>
> > Signed-off-by: Akinobu Mita <akinobu.mita@...il.com>
> > ---
> >  drivers/leds/led-class.c    |  8 ++--
> >  drivers/leds/led-triggers.c | 90 ++++++++++++++++++++++++++++++++++-----------
> >  drivers/leds/leds.h         |  6 +++
> >  include/linux/leds.h        |  5 ---
> >  4 files changed, 79 insertions(+), 30 deletions(-)
> >
> > diff --git a/drivers/leds/led-class.c b/drivers/leds/led-class.c
> > index 4793e77..8b5a1d1 100644
> > --- a/drivers/leds/led-class.c
> > +++ b/drivers/leds/led-class.c
> > @@ -73,13 +73,13 @@ static ssize_t max_brightness_show(struct device *dev,
> >  static DEVICE_ATTR_RO(max_brightness);
> >
> >  #ifdef CONFIG_LEDS_TRIGGERS
> > -static DEVICE_ATTR(trigger, 0644, led_trigger_show, led_trigger_store);
> > -static struct attribute *led_trigger_attrs[] = {
> > -     &dev_attr_trigger.attr,
> > +static BIN_ATTR(trigger, 0644, led_trigger_read, led_trigger_write, 0);
> > +static struct bin_attribute *led_trigger_bin_attrs[] = {
> > +     &bin_attr_trigger,
> >       NULL,
> >  };
> >  static const struct attribute_group led_trigger_group = {
> > -     .attrs = led_trigger_attrs,
> > +     .bin_attrs = led_trigger_bin_attrs,
> >  };
> >  #endif
> >
> > diff --git a/drivers/leds/led-triggers.c b/drivers/leds/led-triggers.c
> > index 8d11a5e..ed5a311 100644
> > --- a/drivers/leds/led-triggers.c
> > +++ b/drivers/leds/led-triggers.c
> > @@ -16,6 +16,7 @@
> >  #include <linux/rwsem.h>
> >  #include <linux/leds.h>
> >  #include <linux/slab.h>
> > +#include <linux/mm.h>
> >  #include "leds.h"
> >
> >  /*
> > @@ -26,9 +27,11 @@ LIST_HEAD(trigger_list);
> >
> >   /* Used by LED Class */
> >
> > -ssize_t led_trigger_store(struct device *dev, struct device_attribute *attr,
> > -             const char *buf, size_t count)
> > +ssize_t led_trigger_write(struct file *filp, struct kobject *kobj,
> > +                       struct bin_attribute *bin_attr, char *buf,
> > +                       loff_t pos, size_t count)
> >  {
> > +     struct device *dev = kobj_to_dev(kobj);
> >       struct led_classdev *led_cdev = dev_get_drvdata(dev);
> >       struct led_trigger *trig;
> >       int ret = count;
> > @@ -64,39 +67,84 @@ ssize_t led_trigger_store(struct device *dev, struct device_attribute *attr,
> >       mutex_unlock(&led_cdev->led_access);
> >       return ret;
> >  }
> > -EXPORT_SYMBOL_GPL(led_trigger_store);
> > +EXPORT_SYMBOL_GPL(led_trigger_write);
> >
> > -ssize_t led_trigger_show(struct device *dev, struct device_attribute *attr,
> > -             char *buf)
> > +__printf(4, 5)
> > +static int led_trigger_snprintf(char *buf, size_t size, bool query,
> > +                             const char *fmt, ...)
> > +{
> > +     va_list args;
> > +     int i;
> > +
> > +     va_start(args, fmt);
> > +     if (query)
> > +             i = vsnprintf(NULL, 0, fmt, args);
> > +     else
> > +             i = vscnprintf(buf, size, fmt, args);
> > +     va_end(args);
> > +
> > +     return i;
> > +}
>
> You only call this in one place, why is it needed like this?  The "old"
> code open-coded this, what is this helping with here?
>
> And what does "query" mean here?  I have no idea how that variable
> matters, or what it does.  Why not just test if buf is NULL or not if
> you don't want to use it?
>
> Ah, you are trying to see how "long" the buffer is going to be.  That
> makes more sense, but just trigger off of the NULL buffer or not, making
> this a bit more "obvious" what you are doing and not tieing two
> parameters to each other (meaning one always reflects the other one).

We cannot simply replace the "query" by checking the buffer is NULL or not.
Because led_trigger_snprintf() is repeatedly called by led_trigger_format()
while increasing 'buf' and decreasing 'size' by the return value of
led_trigger_snprintf() every time.

> > +static int led_trigger_format(char *buf, size_t size, bool query,
> > +                           struct led_classdev *led_cdev)
> >  {
> > -     struct led_classdev *led_cdev = dev_get_drvdata(dev);
> >       struct led_trigger *trig;
> > -     int len = 0;
> > +     int len = led_trigger_snprintf(buf, size, query, "%s",
> > +                                    led_cdev->trigger ? "none" : "[none]");
> > +
> > +     list_for_each_entry(trig, &trigger_list, next_trig) {
> > +             bool hit = led_cdev->trigger &&
> > +                     !strcmp(led_cdev->trigger->name, trig->name);
> > +
> > +             len += led_trigger_snprintf(buf + len, size - len, query,
> > +                                         " %s%s%s", hit ? "[" : "",
> > +                                         trig->name, hit ? "]" : "");
> > +     }
> > +
> > +     len += led_trigger_snprintf(buf + len, size - len, query, "\n");
> > +
> > +     return len;
> > +}
> > +
> > +/*
> > + * It was stupid to create 10000 cpu triggers, but we are stuck with it now.
> > + * Don't make that mistake again. We work around it here by creating binary
> > + * attribute, which is not limited by length. This is _not_ good design, do not
> > + * copy it.
> > + */
> > +ssize_t led_trigger_read(struct file *filp, struct kobject *kobj,
> > +                     struct bin_attribute *attr, char *buf,
> > +                     loff_t pos, size_t count)
> > +{
> > +     struct device *dev = kobj_to_dev(kobj);
> > +     struct led_classdev *led_cdev = dev_get_drvdata(dev);
> > +     void *data;
> > +     int len;
> >
> >       down_read(&triggers_list_lock);
> >       down_read(&led_cdev->trigger_lock);
> >
> > -     if (!led_cdev->trigger)
> > -             len += scnprintf(buf+len, PAGE_SIZE - len, "[none] ");
> > +     len = led_trigger_format(NULL, 0, true, led_cdev);
> > +     data = kvmalloc(len + 1, GFP_KERNEL);
>
> Why kvmalloc() and not just kmalloc()?  How big is this buffer you are
> expecting to have here?

The ledtrig-cpu supports upto 9999 cpus.  If all these cpus were available,
the buffer size would be 78,890 bytes.
(for i in `seq 0 9999`;do echo -n " cpu$i"; done | wc -c)

The intention of this kvmalloc() allocation is to avoid costly allocation
if possible.

> > +     if (data)
> > +             len = led_trigger_format(data, len + 1, false, led_cdev);
> >       else
> > -             len += scnprintf(buf+len, PAGE_SIZE - len, "none ");
> > +             len = -ENOMEM;
> >
> > -     list_for_each_entry(trig, &trigger_list, next_trig) {
> > -             if (led_cdev->trigger && !strcmp(led_cdev->trigger->name,
> > -                                                     trig->name))
> > -                     len += scnprintf(buf+len, PAGE_SIZE - len, "[%s] ",
> > -                                      trig->name);
> > -             else
> > -                     len += scnprintf(buf+len, PAGE_SIZE - len, "%s ",
> > -                                      trig->name);
> > -     }
> >       up_read(&led_cdev->trigger_lock);
> >       up_read(&triggers_list_lock);
>
> Two locks?  Why not 3?  5?  How about just 1?  :)

I don't touch these locks in this patch :)

Locking both locks seems to be necessary to prevent someone from changing
trigger_list or led_cdev->trigger.

> >
> > -     len += scnprintf(len+buf, PAGE_SIZE - len, "\n");
> > +     if (len < 0)
> > +             return len;
>
> You just leaked data if led_trigger_format() returned an error :(
>
> Just return -ENOMEM above if !data, which makes this much simpler.

We are holding the two locks, so we need to release them before return.
Which one do you prefer?

        ...
        data = kvmalloc(len + 1, GFP_KERNEL);
        if (data)
                len = led_trigger_format(data, len + 1, false, led_cdev);
        else
                len = -ENOMEM;

        up_read(&led_cdev->trigger_lock);
        up_read(&triggers_list_lock);

        if (len < 0)
                return len;

vs.

        ...
        data = kvmalloc(len + 1, GFP_KERNEL);
        if (!data) {
                up_read(&led_cdev->trigger_lock);
                up_read(&triggers_list_lock);
                return -ENOMEM;
        }
        len = led_trigger_format(data, len + 1, false, led_cdev);

        up_read(&led_cdev->trigger_lock);
        up_read(&triggers_list_lock);

Powered by blists - more mailing lists