Re: [PATCH v5 4/7] drm/i915/gt: create per-tile sysfs interface
On 13.03.2022 20:45, Andi Shyti wrote:
Hi Andrzej,
I'm sorry, but I'm not fully understanding,
+struct intel_gt *intel_gt_sysfs_get_drvdata(struct device *dev,
+ const char *name)
+{
+ struct kobject *kobj = >kobj;
+
+ /*
+* We are interested at knowing from where the interface
+* has been called, whether it's called from gt/ or from
+* the parent directory.
+* From the interface position it depends also the value of
+* the private data.
+* If the interface is called from gt/ then private data is
+* of the "struct intel_gt *" type, otherwise it's * a
+* "struct drm_i915_private *" type.
+*/
+ if (!is_object_gt(kobj)) {
+ struct drm_i915_private *i915 = kdev_minor_to_i915(dev);
+
+ pr_devel_ratelimited(DEPRECATED
+ "%s (pid %d) is accessing deprecated %s "
+ "sysfs control, please use gt/gt/%s instead\n",
+ current->comm, task_pid_nr(current), name, name);
+ return to_gt(i915);
+ }
+
+ return kobj_to_gt(kobj);
It took some time for me to understand what is going on here.
We have dev argument which sometimes can point to "struct device", sometimes
to "struct kobj_gt", but it's type suggests differently, quite ugly.
I wonder if wouldn't be better to use __ATTR instead of DEVICE_ATTR* as in
case of intel_engines_add_sysfs. This way abstractions would look better,
hopefully.
How would it help?
The difference is that I'm adding twice different interfaces with
the same name and different location (i.e. different object). The
legacy intrefaces inherit the object from drm and I'm preserving
that reference.
While the new objects would derive from the previous and they are
pretty much like intel_engines_add_sysfs().
I was not clear on the issue. Here in case of 'id' attribute it is defined
as device_attribute, but in kobj_type.sysfs_ops you assign formally
incompatible _sysfs_ops.
'kobj_sysfs_ops' is of the type 'kobj_type'.
Yes, but for example kobj_sysfs_ops.show points to function
kobj_attr_show, and kobj_attr_show expects that it's attr argument is
embedded in kobj_attribute[1], but this is not true in case of 'id'
attribute - it is embedded in device_attribute.
In short kobj_sysfs_ops should be used only with attrs embeded in
kobj_attribute, unless I missed sth.
[1]: https://elixir.bootlin.com/linux/latest/source/lib/kobject.c#L836
kobj_sysfs_ops expects kobj_attribute! Fortunately kobj_attribute is 'binary
compatible' with device_attribute and kobj is at beginning of struct device
as well, so it does not blow up, but I wouldn't say it is clean solution :)
If you look at intel_engines_add_sysfs you can see that all attributes are
defined as kobj_attribute.
That's exactly the approach I use in the next patches for the
power management files, I use "struct kobj_gt" wrapped around
"struct kobject". But I'm using that only for the GT files.
But attributes are still defined using DEVICE_ATTR* macros, ie they are
embedded in device_attribute, so the problem is the same - you are using
kobj_sysfs_ops with device_attribute.
Are you, btw, suggesting to use this same approache also for the
legacy files that for now have a pointer to the drm kobject? This
way I would need to add more information, like the pointer to
i915 and gt_id. This way I wouldn't need the files above that
look hacky to you. Is this what you mean?
Positive feedback is more difficult :)
I am little bit lost in possible solutions, after grepping other drivers
I have not good advice about proper handling of such situation, *beside
splitting the interface*.
For sure attrs used in device/power must be embedded in
device_attribute. So if you do not want to split interface, then it
implies GTs attrs must be also in device_attribute. Then maybe creating
custom sysfs_ops would help??? I am not sure.
Regards
Andrzej
Andi
Re: [PATCH v5 4/7] drm/i915/gt: create per-tile sysfs interface
> > > > > +struct intel_gt *intel_gt_sysfs_get_drvdata(struct device *dev,
> > > > > + const char *name)
> > > > > +{
> > > > > + struct kobject *kobj = >kobj;
> > > > > +
> > > > > + /*
> > > > > + * We are interested at knowing from where the interface
> > > > > + * has been called, whether it's called from gt/ or from
> > > > > + * the parent directory.
> > > > > + * From the interface position it depends also the value of
> > > > > + * the private data.
> > > > > + * If the interface is called from gt/ then private data is
> > > > > + * of the "struct intel_gt *" type, otherwise it's * a
> > > > > + * "struct drm_i915_private *" type.
> > > > > + */
> > > > > + if (!is_object_gt(kobj)) {
> > > > > + struct drm_i915_private *i915 = kdev_minor_to_i915(dev);
> > > > > +
> > > > > + pr_devel_ratelimited(DEPRECATED
> > > > > + "%s (pid %d) is accessing deprecated %s "
> > > > > + "sysfs control, please use gt/gt/%s
> > > > > instead\n",
> > > > > + current->comm, task_pid_nr(current), name,
> > > > > name);
> > > > > + return to_gt(i915);
> > > > > + }
> > > > > +
> > > > > + return kobj_to_gt(kobj);
> > > > It took some time for me to understand what is going on here.
> > > > We have dev argument which sometimes can point to "struct device",
> > > > sometimes
> > > > to "struct kobj_gt", but it's type suggests differently, quite ugly.
> > > > I wonder if wouldn't be better to use __ATTR instead of DEVICE_ATTR* as
> > > > in
> > > > case of intel_engines_add_sysfs. This way abstractions would look
> > > > better,
> > > > hopefully.
> > > How would it help?
> > >
> > > The difference is that I'm adding twice different interfaces with
> > > the same name and different location (i.e. different object). The
> > > legacy intrefaces inherit the object from drm and I'm preserving
> > > that reference.
> > >
> > > While the new objects would derive from the previous and they are
> > > pretty much like intel_engines_add_sysfs().
> >
> > I was not clear on the issue. Here in case of 'id' attribute it is defined
> > as device_attribute, but in kobj_type.sysfs_ops you assign formally
> > incompatible _sysfs_ops.
>
> 'kobj_sysfs_ops' is of the type 'kobj_type'.
>
> > kobj_sysfs_ops expects kobj_attribute! Fortunately kobj_attribute is 'binary
> > compatible' with device_attribute and kobj is at beginning of struct device
> > as well, so it does not blow up, but I wouldn't say it is clean solution :)
> > If you look at intel_engines_add_sysfs you can see that all attributes are
> > defined as kobj_attribute.
>
> That's exactly the approach I use in the next patches for the
> power management files, I use "struct kobj_gt" wrapped around
> "struct kobject". But I'm using that only for the GT files.
>
> Are you, btw, suggesting to use this same approache also for the
> legacy files that for now have a pointer to the drm kobject? This
> way I would need to add more information, like the pointer to
> i915 and gt_id. This way I wouldn't need the files above that
> look hacky to you. Is this what you mean?
Still this wouldn't solve it because I am merging the legacy
interfaces to an existing kobject and creating new kobjects for
the new interfaces that go under gt. I would need some other
ugly hack to have things coming around.
Unless I misunderstood you.
Andi
Re: [PATCH v5 4/7] drm/i915/gt: create per-tile sysfs interface
Hi Andrzej,
I'm sorry, but I'm not fully understanding,
> > > > +struct intel_gt *intel_gt_sysfs_get_drvdata(struct device *dev,
> > > > + const char *name)
> > > > +{
> > > > + struct kobject *kobj = >kobj;
> > > > +
> > > > + /*
> > > > +* We are interested at knowing from where the interface
> > > > +* has been called, whether it's called from gt/ or from
> > > > +* the parent directory.
> > > > +* From the interface position it depends also the value of
> > > > +* the private data.
> > > > +* If the interface is called from gt/ then private data is
> > > > +* of the "struct intel_gt *" type, otherwise it's * a
> > > > +* "struct drm_i915_private *" type.
> > > > +*/
> > > > + if (!is_object_gt(kobj)) {
> > > > + struct drm_i915_private *i915 = kdev_minor_to_i915(dev);
> > > > +
> > > > + pr_devel_ratelimited(DEPRECATED
> > > > + "%s (pid %d) is accessing deprecated %s "
> > > > + "sysfs control, please use gt/gt/%s
> > > > instead\n",
> > > > + current->comm, task_pid_nr(current), name,
> > > > name);
> > > > + return to_gt(i915);
> > > > + }
> > > > +
> > > > + return kobj_to_gt(kobj);
> > > It took some time for me to understand what is going on here.
> > > We have dev argument which sometimes can point to "struct device",
> > > sometimes
> > > to "struct kobj_gt", but it's type suggests differently, quite ugly.
> > > I wonder if wouldn't be better to use __ATTR instead of DEVICE_ATTR* as in
> > > case of intel_engines_add_sysfs. This way abstractions would look better,
> > > hopefully.
> > How would it help?
> >
> > The difference is that I'm adding twice different interfaces with
> > the same name and different location (i.e. different object). The
> > legacy intrefaces inherit the object from drm and I'm preserving
> > that reference.
> >
> > While the new objects would derive from the previous and they are
> > pretty much like intel_engines_add_sysfs().
>
> I was not clear on the issue. Here in case of 'id' attribute it is defined
> as device_attribute, but in kobj_type.sysfs_ops you assign formally
> incompatible _sysfs_ops.
'kobj_sysfs_ops' is of the type 'kobj_type'.
> kobj_sysfs_ops expects kobj_attribute! Fortunately kobj_attribute is 'binary
> compatible' with device_attribute and kobj is at beginning of struct device
> as well, so it does not blow up, but I wouldn't say it is clean solution :)
> If you look at intel_engines_add_sysfs you can see that all attributes are
> defined as kobj_attribute.
That's exactly the approach I use in the next patches for the
power management files, I use "struct kobj_gt" wrapped around
"struct kobject". But I'm using that only for the GT files.
Are you, btw, suggesting to use this same approache also for the
legacy files that for now have a pointer to the drm kobject? This
way I would need to add more information, like the pointer to
i915 and gt_id. This way I wouldn't need the files above that
look hacky to you. Is this what you mean?
Andi
Re: [PATCH v5 4/7] drm/i915/gt: create per-tile sysfs interface
On 07.03.2022 00:04, Andi Shyti wrote:
Hi Andrzej,
[...]
+bool is_object_gt(struct kobject *kobj)
+{
+ return !strncmp(kobj->name, "gt", 2);
+}
It looks quite fragile, at the moment I do not have better idea:) maybe
after reviewing the rest of the patches.
yeah... it's not pretty, I agree, but I couldn't come up with a
better way of doing it.
+static struct intel_gt *kobj_to_gt(struct kobject *kobj)
+{
+ return container_of(kobj, struct kobj_gt, base)->gt;
+}
+
+struct intel_gt *intel_gt_sysfs_get_drvdata(struct device *dev,
+ const char *name)
+{
+ struct kobject *kobj = >kobj;
+
+ /*
+* We are interested at knowing from where the interface
+* has been called, whether it's called from gt/ or from
+* the parent directory.
+* From the interface position it depends also the value of
+* the private data.
+* If the interface is called from gt/ then private data is
+* of the "struct intel_gt *" type, otherwise it's * a
+* "struct drm_i915_private *" type.
+*/
+ if (!is_object_gt(kobj)) {
+ struct drm_i915_private *i915 = kdev_minor_to_i915(dev);
+
+ pr_devel_ratelimited(DEPRECATED
+ "%s (pid %d) is accessing deprecated %s "
+ "sysfs control, please use gt/gt/%s instead\n",
+ current->comm, task_pid_nr(current), name, name);
+ return to_gt(i915);
+ }
+
+ return kobj_to_gt(kobj);
It took some time for me to understand what is going on here.
We have dev argument which sometimes can point to "struct device", sometimes
to "struct kobj_gt", but it's type suggests differently, quite ugly.
I wonder if wouldn't be better to use __ATTR instead of DEVICE_ATTR* as in
case of intel_engines_add_sysfs. This way abstractions would look better,
hopefully.
How would it help?
The difference is that I'm adding twice different interfaces with
the same name and different location (i.e. different object). The
legacy intrefaces inherit the object from drm and I'm preserving
that reference.
While the new objects would derive from the previous and they are
pretty much like intel_engines_add_sysfs().
I was not clear on the issue. Here in case of 'id' attribute it is
defined as device_attribute, but in kobj_type.sysfs_ops you assign
formally incompatible _sysfs_ops.
kobj_sysfs_ops expects kobj_attribute! Fortunately kobj_attribute is
'binary compatible' with device_attribute and kobj is at beginning of
struct device as well, so it does not blow up, but I wouldn't say it is
clean solution :)
If you look at intel_engines_add_sysfs you can see that all attributes
are defined as kobj_attribute.
Regards
Andrzej
[...]
+struct kobject *
+intel_gt_create_kobj(struct intel_gt *gt, struct kobject *dir, const char
*name)
+{
+ struct kobj_gt *kg;
+
+ kg = kzalloc(sizeof(*kg), GFP_KERNEL);
+ if (!kg)
+ return NULL;
+
+ kobject_init(>base, _gt_type);
+ kg->gt = gt;
+
+ /* xfer ownership to sysfs tree */
+ if (kobject_add(>base, dir, "%s", name)) {
+ kobject_put(>base);
+ return NULL;
+ }
+
+ return >base; /* borrowed ref */
+}
+
+void intel_gt_sysfs_register(struct intel_gt *gt)
+{
+ struct kobject *dir;
+ char name[80];
+
+ snprintf(name, sizeof(name), "gt%d", gt->info.id);
+
+ dir = intel_gt_create_kobj(gt, gt->i915->sysfs_gt, name);
+ if (!dir) {
+ drm_warn(>i915->drm,
+"failed to initialize %s sysfs root\n", name);
+ return;
+ }
+}
Squashing intel_gt_create_kobj into intel_gt_sysfs_register would simplify
code and allows drop snprintf to local array.
right!
+static struct kobject *i915_setup_gt_sysfs(struct kobject *parent)
+{
+ return kobject_create_and_add("gt", parent);
+}
+
void i915_setup_sysfs(struct drm_i915_private *dev_priv)
{
struct device *kdev = dev_priv->drm.primary->kdev;
@@ -538,6 +543,11 @@ void i915_setup_sysfs(struct drm_i915_private *dev_priv)
if (ret)
drm_err(_priv->drm, "RPS sysfs setup failed\n");
+ dev_priv->sysfs_gt = i915_setup_gt_sysfs(>kobj);
Why not directly kobject_create_and_add("gt", parent) ? up to you.
of course!
[...]
Thanks a lot for the review,
Andi
Re: [PATCH v5 4/7] drm/i915/gt: create per-tile sysfs interface
Hi Andrzej,
[...]
> > +bool is_object_gt(struct kobject *kobj)
> > +{
> > + return !strncmp(kobj->name, "gt", 2);
> > +}
>
> It looks quite fragile, at the moment I do not have better idea:) maybe
> after reviewing the rest of the patches.
yeah... it's not pretty, I agree, but I couldn't come up with a
better way of doing it.
> > +static struct intel_gt *kobj_to_gt(struct kobject *kobj)
> > +{
> > + return container_of(kobj, struct kobj_gt, base)->gt;
> > +}
> > +
> > +struct intel_gt *intel_gt_sysfs_get_drvdata(struct device *dev,
> > + const char *name)
> > +{
> > + struct kobject *kobj = >kobj;
> > +
> > + /*
> > +* We are interested at knowing from where the interface
> > +* has been called, whether it's called from gt/ or from
> > +* the parent directory.
> > +* From the interface position it depends also the value of
> > +* the private data.
> > +* If the interface is called from gt/ then private data is
> > +* of the "struct intel_gt *" type, otherwise it's * a
> > +* "struct drm_i915_private *" type.
> > +*/
> > + if (!is_object_gt(kobj)) {
> > + struct drm_i915_private *i915 = kdev_minor_to_i915(dev);
> > +
> > + pr_devel_ratelimited(DEPRECATED
> > + "%s (pid %d) is accessing deprecated %s "
> > + "sysfs control, please use gt/gt/%s instead\n",
> > + current->comm, task_pid_nr(current), name, name);
> > + return to_gt(i915);
> > + }
> > +
> > + return kobj_to_gt(kobj);
>
> It took some time for me to understand what is going on here.
> We have dev argument which sometimes can point to "struct device", sometimes
> to "struct kobj_gt", but it's type suggests differently, quite ugly.
> I wonder if wouldn't be better to use __ATTR instead of DEVICE_ATTR* as in
> case of intel_engines_add_sysfs. This way abstractions would look better,
> hopefully.
How would it help?
The difference is that I'm adding twice different interfaces with
the same name and different location (i.e. different object). The
legacy intrefaces inherit the object from drm and I'm preserving
that reference.
While the new objects would derive from the previous and they are
pretty much like intel_engines_add_sysfs().
[...]
> > +struct kobject *
> > +intel_gt_create_kobj(struct intel_gt *gt, struct kobject *dir, const char
> > *name)
> > +{
> > + struct kobj_gt *kg;
> > +
> > + kg = kzalloc(sizeof(*kg), GFP_KERNEL);
> > + if (!kg)
> > + return NULL;
> > +
> > + kobject_init(>base, _gt_type);
> > + kg->gt = gt;
> > +
> > + /* xfer ownership to sysfs tree */
> > + if (kobject_add(>base, dir, "%s", name)) {
> > + kobject_put(>base);
> > + return NULL;
> > + }
> > +
> > + return >base; /* borrowed ref */
> > +}
> > +
> > +void intel_gt_sysfs_register(struct intel_gt *gt)
> > +{
> > + struct kobject *dir;
> > + char name[80];
> > +
> > + snprintf(name, sizeof(name), "gt%d", gt->info.id);
> > +
> > + dir = intel_gt_create_kobj(gt, gt->i915->sysfs_gt, name);
> > + if (!dir) {
> > + drm_warn(>i915->drm,
> > +"failed to initialize %s sysfs root\n", name);
> > + return;
> > + }
> > +}
>
> Squashing intel_gt_create_kobj into intel_gt_sysfs_register would simplify
> code and allows drop snprintf to local array.
right!
> > +static struct kobject *i915_setup_gt_sysfs(struct kobject *parent)
> > +{
> > + return kobject_create_and_add("gt", parent);
> > +}
> > +
> > void i915_setup_sysfs(struct drm_i915_private *dev_priv)
> > {
> > struct device *kdev = dev_priv->drm.primary->kdev;
> > @@ -538,6 +543,11 @@ void i915_setup_sysfs(struct drm_i915_private
> > *dev_priv)
> > if (ret)
> > drm_err(_priv->drm, "RPS sysfs setup failed\n");
> > + dev_priv->sysfs_gt = i915_setup_gt_sysfs(>kobj);
>
> Why not directly kobject_create_and_add("gt", parent) ? up to you.
of course!
[...]
Thanks a lot for the review,
Andi
Re: [PATCH v5 4/7] drm/i915/gt: create per-tile sysfs interface
On 17.02.2022 15:41, Andi Shyti wrote:
Now that we have tiles we want each of them to have its own
interface. A directory "gt/" is created under "cardN/" that will
contain as many diroctories as the tiles.
In the coming patches tile related interfaces will be added. For
now the sysfs gt structure simply has an id interface related
to the current tile count.
The directory structure will follow this scheme:
/sys/.../card0
└── gt
├── gt0
│ └── id
:
:
└─- gtN
└── id
This new set of interfaces will be a basic tool for system
managers and administrators when using i915.
Signed-off-by: Andi Shyti
Cc: Matt Roper
Cc: Sujaritha Sundaresan
Cc: Tvrtko Ursulin
Reviewed-by: Sujaritha Sundaresan
---
drivers/gpu/drm/i915/Makefile| 1 +
drivers/gpu/drm/i915/gt/intel_gt.c | 2 +
drivers/gpu/drm/i915/gt/intel_gt_sysfs.c | 118 +++
drivers/gpu/drm/i915/gt/intel_gt_sysfs.h | 34 +++
drivers/gpu/drm/i915/i915_drv.h | 2 +
drivers/gpu/drm/i915/i915_sysfs.c| 12 ++-
drivers/gpu/drm/i915/i915_sysfs.h| 3 +
7 files changed, 171 insertions(+), 1 deletion(-)
create mode 100644 drivers/gpu/drm/i915/gt/intel_gt_sysfs.c
create mode 100644 drivers/gpu/drm/i915/gt/intel_gt_sysfs.h
diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile
index 9d588d936e3d..277064b00afd 100644
--- a/drivers/gpu/drm/i915/Makefile
+++ b/drivers/gpu/drm/i915/Makefile
@@ -105,6 +105,7 @@ gt-y += \
gt/intel_gt_pm_debugfs.o \
gt/intel_gt_pm_irq.o \
gt/intel_gt_requests.o \
+ gt/intel_gt_sysfs.o \
gt/intel_gtt.o \
gt/intel_llc.o \
gt/intel_lrc.o \
diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c
b/drivers/gpu/drm/i915/gt/intel_gt.c
index 8c64b81e9ec9..0f080bbad043 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt.c
+++ b/drivers/gpu/drm/i915/gt/intel_gt.c
@@ -26,6 +26,7 @@
#include "intel_rc6.h"
#include "intel_renderstate.h"
#include "intel_rps.h"
+#include "intel_gt_sysfs.h"
#include "intel_uncore.h"
#include "shmem_utils.h"
@@ -458,6 +459,7 @@ void intel_gt_driver_register(struct intel_gt *gt)
intel_rps_driver_register(>rps);
intel_gt_debugfs_register(gt);
+ intel_gt_sysfs_register(gt);
}
static int intel_gt_init_scratch(struct intel_gt *gt, unsigned int size)
diff --git a/drivers/gpu/drm/i915/gt/intel_gt_sysfs.c
b/drivers/gpu/drm/i915/gt/intel_gt_sysfs.c
new file mode 100644
index ..0206e9aa4867
--- /dev/null
+++ b/drivers/gpu/drm/i915/gt/intel_gt_sysfs.c
@@ -0,0 +1,118 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2022 Intel Corporation
+ */
+
+#include
+#include
+#include
+#include
+#include
+
+#include "i915_drv.h"
+#include "i915_sysfs.h"
+#include "intel_gt.h"
+#include "intel_gt_sysfs.h"
+#include "intel_gt_types.h"
+#include "intel_rc6.h"
+
+bool is_object_gt(struct kobject *kobj)
+{
+ return !strncmp(kobj->name, "gt", 2);
+}
It looks quite fragile, at the moment I do not have better idea:) maybe
after reviewing the rest of the patches.
+
+static struct intel_gt *kobj_to_gt(struct kobject *kobj)
+{
+ return container_of(kobj, struct kobj_gt, base)->gt;
+}
+
+struct intel_gt *intel_gt_sysfs_get_drvdata(struct device *dev,
+ const char *name)
+{
+ struct kobject *kobj = >kobj;
+
+ /*
+* We are interested at knowing from where the interface
+* has been called, whether it's called from gt/ or from
+* the parent directory.
+* From the interface position it depends also the value of
+* the private data.
+* If the interface is called from gt/ then private data is
+* of the "struct intel_gt *" type, otherwise it's * a
+* "struct drm_i915_private *" type.
+*/
+ if (!is_object_gt(kobj)) {
+ struct drm_i915_private *i915 = kdev_minor_to_i915(dev);
+
+ pr_devel_ratelimited(DEPRECATED
+ "%s (pid %d) is accessing deprecated %s "
+ "sysfs control, please use gt/gt/%s instead\n",
+ current->comm, task_pid_nr(current), name, name);
+ return to_gt(i915);
+ }
+
+ return kobj_to_gt(kobj);
It took some time for me to understand what is going on here.
We have dev argument which sometimes can point to "struct device",
sometimes to "struct kobj_gt", but it's type suggests differently, quite
ugly.
I wonder if wouldn't be better to use __ATTR instead of DEVICE_ATTR* as
in case of intel_engines_add_sysfs. This way abstractions would look
better, hopefully.
+}
+
+static ssize_t id_show(struct device *dev,
+ struct device_attribute *attr,
+ char *buf)
+{
+ struct intel_gt *gt =
