[PATCH v3 46/57] perf: Simplify pmu_dev_alloc()

[Greg KH]

On Mon, Jun 12, 2023 at 02:18:03PM +0200, Greg KH wrote:
> On Mon, Jun 12, 2023 at 11:44:00AM +0200, Peter Zijlstra wrote:
> > On Mon, Jun 12, 2023 at 11:07:59AM +0200, Peter Zijlstra wrote:
> > > 
> > > Signed-off-by: Peter Zijlstra (Intel) <peterz at infradead.org>
> > > ---
> > >  kernel/events/core.c |   65 ++++++++++++++++++++++++---------------------------
> > >  1 file changed, 31 insertions(+), 34 deletions(-)
> > > 
> > > --- a/kernel/events/core.c
> > > +++ b/kernel/events/core.c
> > > @@ -11285,49 +11285,46 @@ static void pmu_dev_release(struct devic
> > >  
> > >  static int pmu_dev_alloc(struct pmu *pmu)
> > >  {
> > > +	int ret;
> > >  
> > > +	struct device *dev __free(put_device) =
> > > +		kzalloc(sizeof(struct device), GFP_KERNEL);
> > > +	if (!dev)
> > > +		return -ENOMEM;
> > >  
> > > +	dev->groups = pmu->attr_groups;
> > > +	device_initialize(dev);
> > >  
> > > +	dev_set_drvdata(dev, pmu);
> > > +	dev->bus = &pmu_bus;
> > > +	dev->release = pmu_dev_release;
> > >  
> > > +	ret = dev_set_name(dev, "%s", pmu->name);
> > >  	if (ret)
> > > +		return ret;
> > >  
> > > +	ret = device_add(dev);
> > >  	if (ret)
> > > +		return ret;
> > >  
> > > +	struct device *del __free(device_del) = dev;
> > 
> > Greg, I'm not much familiar with the whole device model, but it seems
> > unfortunate to me that one has to call device_del() explicitly if we
> > already have a put_device() queued.
> > 
> > Is there a saner way to write this?
> 
> Ok, the "problem" here is that you have decided to do the "complex" way
> to initialize a struct device.  And as such, you have to do more
> housekeeping than if you were to just use the simple interface.
> 
> The rule is, after you call device_initialize() you HAVE to call
> put_device() on the pointer if something goes wrong and you want to
> clean up properly.  Unless you have called device_add(), and at that
> point in time, then you HAVE to call device_del() if the device_add()
> call succeeded.  If the device_add() call failed, then you HAVE to call
> put_device().
> 
> Yeah, it's a pain, but you are trying to hand-roll code that is not a
> "normal" path for a struct device, sorry.
> 
> I don't know if you really can encode all of that crazy logic in the
> cleanup api, UNLESS you can "switch" the cleanup function at a point in
> time (i.e. after device_add() is successful).  Is that possible?
> 
> Anyway, let me see about just cleaning up this code in general, I don't
> think you need the complex interface here for a tiny struct device at
> all, which would make this specific instance moot :)
> 
> Also, nit, you are racing with userspace by attempting to add new device
> files _AFTER_ the device is registered with the driver core, this whole
> thing can be made more simpler I hope, give me a bit...

Nope, I was wrong, I can fix the race condition, but the logic here for
how to init and clean up on errors is right, and you want this because
you are a bus and so, you need the two-step init/teardown process,
sorry.

Here's the patch I came up with to get rid of the race, but doesn't
really help you out here at all :(

------------------------
>From foo at baz Mon Jun 12 03:07:54 PM CEST 2023
Date: Mon, 12 Jun 2023 15:07:54 +0200
To: Greg KH <gregkh at linuxfoundation.org>
From: Greg Kroah-Hartman <gregkh at linuxfoundation.org>
Subject: [PATCH] perf/core: fix narrow startup race when creating the perf nr_addr_filters sysfs file


Signed-off-by: Greg Kroah-Hartman <gregkh at linuxfoundation.org>


diff --git a/kernel/events/core.c b/kernel/events/core.c
index db016e418931..d2a6182ad090 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -11351,9 +11351,32 @@ static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
 static struct attribute *pmu_dev_attrs[] = {
 	&dev_attr_type.attr,
 	&dev_attr_perf_event_mux_interval_ms.attr,
+	&dev_attr_nr_addr_filters.attr,
+	NULL,
+};
+
+static umode_t pmu_dev_is_visible(struct kobject *kobj, struct attribute *a, int n)
+{
+	struct device *dev = kobj_to_dev(kobj);
+	struct pmu *pmu = dev_get_drvdata(dev);
+
+	if (!pmu->nr_addr_filters)
+		return 0;
+
+	return a->mode;
+
+	return 0;
+}
+
+static struct attribute_group pmu_dev_attr_group = {
+	.is_visible = pmu_dev_is_visible,
+	.attrs = pmu_dev_attrs,
+};
+
+const static struct attribute_group *pmu_dev_groups[] = {
+	&pmu_dev_attr_group,
 	NULL,
 };
-ATTRIBUTE_GROUPS(pmu_dev);
 
 static int pmu_bus_running;
 static struct bus_type pmu_bus = {
@@ -11389,18 +11412,11 @@ static int pmu_dev_alloc(struct pmu *pmu)
 	if (ret)
 		goto free_dev;
 
-	/* For PMUs with address filters, throw in an extra attribute: */
-	if (pmu->nr_addr_filters)
-		ret = device_create_file(pmu->dev, &dev_attr_nr_addr_filters);
-
-	if (ret)
-		goto del_dev;
-
-	if (pmu->attr_update)
+	if (pmu->attr_update) {
 		ret = sysfs_update_groups(&pmu->dev->kobj, pmu->attr_update);
-
-	if (ret)
-		goto del_dev;
+		if (ret)
+			goto del_dev;
+	}
 
 out:
 	return ret;