Skip to contentSkip to conversation

system hang on ILK/SNB/IVB

REGRESSION Superseded Gabriel Feceoru 8 participants 25 messages Mar 30, 2016 – Apr 1, 2016

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

This commit causes a hang while running kms suspend tests (kms_pipe_crc_basic@suspend-read-crc-pipe-*) on ILK/SNB/IVB, affecting CI.

Probably the same problem with the one in v2, but on older HW.

commit a7442b93cf32c1e1ddb721a26cd1f92302e2a222

Author: Lukas Wunner <[email protected]>
Date:   Wed Mar 9 12:52:53 2016 +0100

drm/i915: Fix races on fbdev

The ->lastclose callback invokes intel_fbdev_restore_mode() and has been witnessed to run before intel_fbdev_initial_config_async() has finished.

We might likewise receive hotplug events before we've had a chance to fully set up the fbdev.

Fix by waiting for the asynchronous thread to finish.

v2:
An async_synchronize_full() was also added to intel_fbdev_set_suspend() in v1 which turned out to be entirely gratuitous. It caused a deadlock on suspend (discovered by CI, thanks to Damien Lespiau and Tomi Sarvela for CI support) and was unnecessary since a device is never suspended until its ->probe callback (and all asynchronous tasks it scheduled) have finished. See dpm_prepare(), which calls wait_for_device_probe(), which calls async_synchronize_full().

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

Link:
http://patchwork.freedesktop.org/patch/msgid/[email protected]

Regards,
Gabriel
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#1 lore raw

PATCH 1/1 drm/i915: Protect fbdev across slow or failed initialisation

If the initialisation fails, we may be left with a dangling pointer with an incomplete fbdev structure. Here we want to disable internal calls into fbdev. Similarly, the initialisation may be slow and we haven't yet enabled the fbdev (e.g. quick suspend or last-close before the async init completes).

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

---
 drivers/gpu/drm/i915/intel_fbdev.c | 48 ++++++++++++++++++++++++--------------
 1 file changed, 30 insertions(+), 18 deletions(-)
drivers/gpu/drm/i915/intel_fbdev.c +30 -18
index 51a5e39e52f2..0bae91268a12 100644
@@ -114,6 +114,24 @@ static struct fb_ops intelfb_ops = {
 	.fb_debug_leave = drm_fb_helper_debug_leave,
 };
 
+static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
+{
+	struct drm_i915_device *dev_priv = to_i915(dev);
+	struct fb_info *info;
+
+	if (dev_priv->fbdev == NULL)
+		return NULL;
+
+	info = ifbdev->helper.fbdev;
+	if (info->screen_base == NULL)
+		return NULL;
+
+	if (info->state != FBINFO_STATE_RUNNING)
+		return NULL;
+
+	return dev_priv->fbdev;
+}
+
 static int intelfb_alloc(struct drm_fb_helper *helper,
 			 struct drm_fb_helper_surface_size *sizes)
 {
@@ -764,6 +782,8 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
 		return;
 
 	info = ifbdev->helper.fbdev;
+	if (info->screen_base == NULL)
+		return;
 
 	if (synchronous) {
 		/* Flush any pending work to turn the console on, and then
@@ -805,32 +825,24 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
 
 void intel_fbdev_output_poll_changed(struct drm_device *dev)
 {
-	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
+
+	if (ifbdev == NULL)
+		return;
 
-	async_synchronize_full();
-	if (dev_priv->fbdev)
-		drm_fb_helper_hotplug_event(&dev_priv->fbdev->helper);
+	drm_fb_helper_hotplug_event(&ifbdev->helper);
 }
 
 void intel_fbdev_restore_mode(struct drm_device *dev)
 {
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
 
-	async_synchronize_full();
-	if (!ifbdev)
+	if (ifbdev == NULL)
 		return;
 
-	fb_helper = &ifbdev->helper;
-
-	ret = drm_fb_helper_restore_fbdev_mode_unlocked(fb_helper);
-	if (ret) {
-		DRM_DEBUG("failed to restore crtc mode\n");
-	} else {
-		mutex_lock(&fb_helper->dev->struct_mutex);
+	if (drm_fb_helper_restore_fbdev_mode_unlocked(&ifbdev->helper) == 0) {
+		mutex_lock(&dev->struct_mutex);
 		intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);
-		mutex_unlock(&fb_helper->dev->struct_mutex);
+		mutex_unlock(&dev->struct_mutex);
 	}
 }
Signature · 6 lines
2.8.0.rc3

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw

Re: [PATCH] drm/i915: Protect fbdev across slow or failed initialisation

reproduce:

Hi Chris,

[auto build test ERROR on drm-intel/for-linux-next]
[cannot apply to v4.6-rc1 next-20160330]
[if your patch is applied to the wrong git tree, please drop us a note to help improving the system]

url:    https://github.com/0day-ci/linux/commits/Chris-Wilson/drm-i915-Protect-fbdev-across-slow-or-failed-initialisation/20160331-015912
base:   git://anongit.freedesktop.org/drm-intel for-linux-next
config: x86_64-randconfig-x012-03310059 (attached as .config)
reproduce:
        # save the attached .config to linux build tree
        make ARCH=x86_64 

All error/warnings (new ones prefixed by >>):

drivers/gpu/drm/i915/intel_fbdev.c: In function 'intel_fbdev_get_if_active':

>> drivers/gpu/drm/i915/intel_fbdev.c:119:37: warning: initialization from incompatible pointer type [-Wincompatible-pointer-types]
     struct drm_i915_device *dev_priv = to_i915(dev);
                                        ^
>> drivers/gpu/drm/i915/intel_fbdev.c:122:14: error: dereferencing pointer to incomplete type 'struct drm_i915_device'
     if (dev_priv->fbdev == NULL)
                 ^
>> drivers/gpu/drm/i915/intel_fbdev.c:125:9: error: 'ifbdev' undeclared (first use in this function)
     info = ifbdev->helper.fbdev;
            ^
   drivers/gpu/drm/i915/intel_fbdev.c:125:9: note: each undeclared identifier is reported only once for each function it appears in
>> drivers/gpu/drm/i915/intel_fbdev.c:133:1: warning: control reaches end of non-void function [-Wreturn-type]
    }
    ^

vim +122 drivers/gpu/drm/i915/intel_fbdev.c

113 .fb_debug_enter = drm_fb_helper_debug_enter,
114 .fb_debug_leave = drm_fb_helper_debug_leave,
115 };
116
117 static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
118 {

 > 119		struct drm_i915_device *dev_priv = to_i915(dev);
   120		struct fb_info *info;
   121	
 > 122		if (dev_priv->fbdev == NULL)
   123			return NULL;
   124	
 > 125		info = ifbdev->helper.fbdev;
   126		if (info->screen_base == NULL)
   127			return NULL;
   128	
   129		if (info->state != FBINFO_STATE_RUNNING)
   130			return NULL;
   131	
   132		return dev_priv->fbdev;
 > 133	}

134
135 static int intelfb_alloc(struct drm_fb_helper *helper,
136 struct drm_fb_helper_surface_size *sizes)

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#3 lore raw

Re: [PATCH] drm/i915: Protect fbdev across slow or failed initialisation

In reply to Chris Wilson
reproduce:

Hi Chris,

[auto build test WARNING on drm-intel/for-linux-next]
[cannot apply to v4.6-rc1 next-20160330]
[if your patch is applied to the wrong git tree, please drop us a note to help improving the system]

url:    https://github.com/0day-ci/linux/commits/Chris-Wilson/drm-i915-Protect-fbdev-across-slow-or-failed-initialisation/20160331-015912
base:   git://anongit.freedesktop.org/drm-intel for-linux-next
config: x86_64-randconfig-x015-03310059 (attached as .config)
reproduce:
        # save the attached .config to linux build tree
        make ARCH=x86_64 

All warnings (new ones prefixed by >>):

drivers/gpu/drm/i915/intel_fbdev.c: In function 'intel_fbdev_get_if_active':
drivers/gpu/drm/i915/intel_fbdev.c:119:37: warning: initialization from incompatible pointer type [-Wincompatible-pointer-types]

     struct drm_i915_device *dev_priv = to_i915(dev);
                                        ^
   In file included from include/uapi/linux/stddef.h:1:0,
                    from include/linux/stddef.h:4,
                    from include/uapi/linux/posix_types.h:4,
                    from include/uapi/linux/types.h:13,
                    from include/linux/types.h:5,
                    from include/linux/async.h:15,
                    from drivers/gpu/drm/i915/intel_fbdev.c:27:
   drivers/gpu/drm/i915/intel_fbdev.c:122:14: error: dereferencing pointer to incomplete type 'struct drm_i915_device'
     if (dev_priv->fbdev == NULL)
                 ^
   include/linux/compiler.h:147:30: note: in definition of macro '__trace_if'
     if (__builtin_constant_p(!!(cond)) ? !!(cond) :   \
                                 ^
>> drivers/gpu/drm/i915/intel_fbdev.c:122:2: note: in expansion of macro 'if'
     if (dev_priv->fbdev == NULL)
     ^
   drivers/gpu/drm/i915/intel_fbdev.c:125:9: error: 'ifbdev' undeclared (first use in this function)
     info = ifbdev->helper.fbdev;
            ^
   drivers/gpu/drm/i915/intel_fbdev.c:125:9: note: each undeclared identifier is reported only once for each function it appears in
   drivers/gpu/drm/i915/intel_fbdev.c:133:1: warning: control reaches end of non-void function [-Wreturn-type]
    }
    ^

vim +/if +122 drivers/gpu/drm/i915/intel_fbdev.c

    21	 * DEALINGS IN THE SOFTWARE.
    22	 *
    23	 * Authors:
    24	 *     David Airlie
    25	 */
    26	
  > 27	#include <linux/async.h>
    28	#include <linux/module.h>
    29	#include <linux/kernel.h>
    30	#include <linux/console.h>
    31	#include <linux/errno.h>
    32	#include <linux/string.h>
    33	#include <linux/mm.h>
    34	#include <linux/tty.h>
    35	#include <linux/sysrq.h>
    36	#include <linux/delay.h>
    37	#include <linux/fb.h>
    38	#include <linux/init.h>
    39	#include <linux/vga_switcheroo.h>
    40	
    41	#include <drm/drmP.h>
    42	#include <drm/drm_crtc.h>
    43	#include <drm/drm_fb_helper.h>
    44	#include "intel_drv.h"
    45	#include <drm/i915_drm.h>
    46	#include "i915_drv.h"
    47	
    48	static int intel_fbdev_set_par(struct fb_info *info)
    49	{
    50		struct drm_fb_helper *fb_helper = info->par;
    51		struct intel_fbdev *ifbdev =
    52			container_of(fb_helper, struct intel_fbdev, helper);
    53		int ret;
    54	
    55		ret = drm_fb_helper_set_par(info);
    56	
    57		if (ret == 0) {
    58			mutex_lock(&fb_helper->dev->struct_mutex);
    59			intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);
    60			mutex_unlock(&fb_helper->dev->struct_mutex);
    61		}
    62	
    63		return ret;
    64	}
    65	
    66	static int intel_fbdev_blank(int blank, struct fb_info *info)
    67	{
    68		struct drm_fb_helper *fb_helper = info->par;
    69		struct intel_fbdev *ifbdev =
    70			container_of(fb_helper, struct intel_fbdev, helper);
    71		int ret;
    72	
    73		ret = drm_fb_helper_blank(blank, info);
    74	
    75		if (ret == 0) {
    76			mutex_lock(&fb_helper->dev->struct_mutex);
    77			intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);
    78			mutex_unlock(&fb_helper->dev->struct_mutex);
    79		}
    80	
    81		return ret;
    82	}
    83	
    84	static int intel_fbdev_pan_display(struct fb_var_screeninfo *var,
    85					   struct fb_info *info)
    86	{
    87		struct drm_fb_helper *fb_helper = info->par;
    88		struct intel_fbdev *ifbdev =
    89			container_of(fb_helper, struct intel_fbdev, helper);
    90	
    91		int ret;
    92		ret = drm_fb_helper_pan_display(var, info);
    93	
    94		if (ret == 0) {
    95			mutex_lock(&fb_helper->dev->struct_mutex);
    96			intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);
    97			mutex_unlock(&fb_helper->dev->struct_mutex);
    98		}
    99	
   100		return ret;
   101	}
   102	
   103	static struct fb_ops intelfb_ops = {
   104		.owner = THIS_MODULE,
   105		.fb_check_var = drm_fb_helper_check_var,
   106		.fb_set_par = intel_fbdev_set_par,
   107		.fb_fillrect = drm_fb_helper_cfb_fillrect,
   108		.fb_copyarea = drm_fb_helper_cfb_copyarea,
   109		.fb_imageblit = drm_fb_helper_cfb_imageblit,
   110		.fb_pan_display = intel_fbdev_pan_display,
   111		.fb_blank = intel_fbdev_blank,
   112		.fb_setcmap = drm_fb_helper_setcmap,
   113		.fb_debug_enter = drm_fb_helper_debug_enter,
   114		.fb_debug_leave = drm_fb_helper_debug_leave,
   115	};
   116	
   117	static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
   118	{
   119		struct drm_i915_device *dev_priv = to_i915(dev);
   120		struct fb_info *info;
   121	
 > 122		if (dev_priv->fbdev == NULL)

123 return NULL;
124
125 info = ifbdev->helper.fbdev;

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#4 lore raw

Re: [PATCH] drm/i915: Protect fbdev across slow or failed initialisation

In reply to Chris Wilson
reproduce:

Hi Chris,

[auto build test ERROR on drm-intel/for-linux-next]
[cannot apply to v4.6-rc1 next-20160330]
[if your patch is applied to the wrong git tree, please drop us a note to help improving the system]

url:    https://github.com/0day-ci/linux/commits/Chris-Wilson/drm-i915-Protect-fbdev-across-slow-or-failed-initialisation/20160331-015912
base:   git://anongit.freedesktop.org/drm-intel for-linux-next
config: x86_64-rhel (attached as .config)
reproduce:
        # save the attached .config to linux build tree
        make ARCH=x86_64 

All errors (new ones prefixed by >>):

drivers/gpu/drm/i915/intel_fbdev.c: In function 'intel_fbdev_get_if_active':
drivers/gpu/drm/i915/intel_fbdev.c:119:37: warning: initialization from incompatible pointer type

     struct drm_i915_device *dev_priv = to_i915(dev);
                                        ^
>> drivers/gpu/drm/i915/intel_fbdev.c:122:14: error: dereferencing pointer to incomplete type
     if (dev_priv->fbdev == NULL)
                 ^
   drivers/gpu/drm/i915/intel_fbdev.c:125:9: error: 'ifbdev' undeclared (first use in this function)
     info = ifbdev->helper.fbdev;
            ^
   drivers/gpu/drm/i915/intel_fbdev.c:125:9: note: each undeclared identifier is reported only once for each function it appears in
   drivers/gpu/drm/i915/intel_fbdev.c:132:17: error: dereferencing pointer to incomplete type
     return dev_priv->fbdev;
                    ^
   drivers/gpu/drm/i915/intel_fbdev.c:133:1: warning: control reaches end of non-void function [-Wreturn-type]
    }
    ^

vim +122 drivers/gpu/drm/i915/intel_fbdev.c

113 .fb_debug_enter = drm_fb_helper_debug_enter,
114 .fb_debug_leave = drm_fb_helper_debug_leave,
115 };
116
117 static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
118 {

 > 119		struct drm_i915_device *dev_priv = to_i915(dev);
   120		struct fb_info *info;
   121	
 > 122		if (dev_priv->fbdev == NULL)

123 return NULL;
124
125 info = ifbdev->helper.fbdev;

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#5 lore raw

Re: [PATCH] drm/i915: Protect fbdev across slow or failed initialisation

In reply to Chris Wilson
Quoted 18 lines · Chris Wilson

If the initialisation fails, we may be left with a dangling pointer with an incomplete fbdev structure. Here we want to disable internal calls into fbdev. Similarly, the initialisation may be slow and we haven't yet enabled the fbdev (e.g. quick suspend or last-close before the async init completes).

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

---
 drivers/gpu/drm/i915/intel_fbdev.c | 48 ++++++++++++++++++++++++--------------
 1 file changed, 30 insertions(+), 18 deletions(-)
drivers/gpu/drm/i915/intel_fbdev.c +0 -0
index 51a5e39e52f2..0bae91268a12 100644

drivers/gpu/drm/i915/intel_fbdev.c static struct fb_ops intelfb_ops = { · line 119

Quoted 6 lines
drivers/gpu/drm/i915/intel_fbdev.c +3 -0
@@ -114,6 +114,24 @@ static struct fb_ops intelfb_ops = {
 	.fb_debug_leave = drm_fb_helper_debug_leave,
 };
 
+static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
+{
+	struct drm_i915_device *dev_priv = to_i915(dev);

As kbuild reported, the danger of retyping a patch from email and forgetting you compiled on a machine without fbdev. s/drm_i915_device/drm_i915_private/
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#17 lore raw
In reply to Gabriel Feceoru
Quoted 4 lines · Gabriel Feceoru

This commit causes a hang while running kms suspend tests (kms_pipe_crc_basic@suspend-read-crc-pipe-*) on ILK/SNB/IVB, affecting CI.

Probably the same problem with the one in v2, but on older HW.

I did check the patchwork/BAT-CI results and it looked clean. Is this a new machine? Should I just revert for now until we have a proper fix? -Daniel

Quoted 38 lines

commit a7442b93cf32c1e1ddb721a26cd1f92302e2a222

Author: Lukas Wunner <[email protected]>
Date:   Wed Mar 9 12:52:53 2016 +0100

drm/i915: Fix races on fbdev

The ->lastclose callback invokes intel_fbdev_restore_mode() and has been witnessed to run before intel_fbdev_initial_config_async() has finished.

We might likewise receive hotplug events before we've had a chance to fully set up the fbdev.

Fix by waiting for the asynchronous thread to finish.

v2:
An async_synchronize_full() was also added to intel_fbdev_set_suspend() in v1 which turned out to be entirely gratuitous. It caused a deadlock on suspend (discovered by CI, thanks to Damien Lespiau and Tomi Sarvela for CI support) and was unnecessary since a device is never suspended until its ->probe callback (and all asynchronous tasks it scheduled) have finished. See dpm_prepare(), which calls wait_for_device_probe(), which calls async_synchronize_full().

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

Regards,
Gabriel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#6 lore raw

PATCH v3 1/1 drm/i915: Protect fbdev across slow or failed initialisation

In reply to Chris Wilson

If the initialisation fails, we may be left with a dangling pointer with an incomplete fbdev structure. Here we want to disable internal calls into fbdev. Similarly, the initialisation may be slow and we haven't yet enabled the fbdev (e.g. quick suspend or last-close before the async init completes).

v3: To create a typo introduced when retyping

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

---
 drivers/gpu/drm/i915/intel_fbdev.c | 48 ++++++++++++++++++++++++--------------
 1 file changed, 30 insertions(+), 18 deletions(-)
drivers/gpu/drm/i915/intel_fbdev.c +30 -18
index 153ea7a3fcf6..5029f927fe0d 100644
@@ -114,6 +114,24 @@ static struct fb_ops intelfb_ops = {
 	.fb_debug_leave = drm_fb_helper_debug_leave,
 };
 
+static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
+{
+	struct drm_i915_private *dev_priv = to_i915(dev);
+	struct fb_info *info;
+
+	if (dev_priv->fbdev == NULL)
+		return NULL;
+
+	info = dev_priv->fbdev->helper.fbdev;
+	if (info->screen_base == NULL)
+		return NULL;
+
+	if (info->state != FBINFO_STATE_RUNNING)
+		return NULL;
+
+	return dev_priv->fbdev;
+}
+
 static int intelfb_alloc(struct drm_fb_helper *helper,
 			 struct drm_fb_helper_surface_size *sizes)
 {
@@ -766,6 +784,8 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
 		return;
 
 	info = ifbdev->helper.fbdev;
+	if (info->screen_base == NULL)
+		return;
 
 	if (synchronous) {
 		/* Flush any pending work to turn the console on, and then
@@ -807,32 +827,24 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
 
 void intel_fbdev_output_poll_changed(struct drm_device *dev)
 {
-	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
+
+	if (ifbdev == NULL)
+		return;
 
-	async_synchronize_full();
-	if (dev_priv->fbdev)
-		drm_fb_helper_hotplug_event(&dev_priv->fbdev->helper);
+	drm_fb_helper_hotplug_event(&ifbdev->helper);
 }
 
 void intel_fbdev_restore_mode(struct drm_device *dev)
 {
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
 
-	async_synchronize_full();
-	if (!ifbdev)
+	if (ifbdev == NULL)
 		return;
 
-	fb_helper = &ifbdev->helper;
-
-	ret = drm_fb_helper_restore_fbdev_mode_unlocked(fb_helper);
-	if (ret) {
-		DRM_DEBUG("failed to restore crtc mode\n");
-	} else {
-		mutex_lock(&fb_helper->dev->struct_mutex);
+	if (drm_fb_helper_restore_fbdev_mode_unlocked(&ifbdev->helper) == 0) {
+		mutex_lock(&dev->struct_mutex);
 		intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);
-		mutex_unlock(&fb_helper->dev->struct_mutex);
+		mutex_unlock(&dev->struct_mutex);
 	}
 }
Signature · 6 lines
2.8.0.rc3

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#18 lore raw
In reply to Gabriel Feceoru

Hi Gabriel,

Quoted 2 lines · Gabriel Feceoru

This commit causes a hang while running kms suspend tests (kms_pipe_crc_basic@suspend-read-crc-pipe-*) on ILK/SNB/IVB, affecting CI.

This happened with v1 but not with v2 of the patch. Please check if somehow v1 ended up in your tree.

v2 passed CI fine, save for one warning not caused by the patch:
https://patchwork.freedesktop.org/series/4068/

For comparison, this was v1:
https://patchwork.freedesktop.org/patch/75840/

Best regards,

Lukas

Quoted 38 lines

Probably the same problem with the one in v2, but on older HW.

commit a7442b93cf32c1e1ddb721a26cd1f92302e2a222

Author: Lukas Wunner <[email protected]>
Date:   Wed Mar 9 12:52:53 2016 +0100

drm/i915: Fix races on fbdev

The ->lastclose callback invokes intel_fbdev_restore_mode() and has been witnessed to run before intel_fbdev_initial_config_async() has finished.

We might likewise receive hotplug events before we've had a chance to fully set up the fbdev.

Fix by waiting for the asynchronous thread to finish.

v2:
An async_synchronize_full() was also added to intel_fbdev_set_suspend() in v1 which turned out to be entirely gratuitous. It caused a deadlock on suspend (discovered by CI, thanks to Damien Lespiau and Tomi Sarvela for CI support) and was unnecessary since a device is never suspended until its ->probe callback (and all asynchronous tasks it scheduled) have finished. See dpm_prepare(), which calls wait_for_device_probe(), which calls async_synchronize_full().

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

Quoted 2 lines

Regards,
Gabriel

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#19 lore raw

Hello Lukas,

The problem with the results in your link is that there is no HSW, ILK, IVB or SNB results. This might give the impression that everything is well.

Most damning is lack of HSW-gt2 and SNB-dellxps: those machines hang on to APC, and have run quite stably for every Patchwork run. The case isn't strong enough yet that series should fail if either of those won't run, but it might be so in future.

Tomi

Quoted 12 lines · On Wednesday 30 March 2016 23:35:08 Lukas Wunner wrote:

Hi Gabriel,

Quoted 2 lines · Gabriel Feceoru

This commit causes a hang while running kms suspend tests (kms_pipe_crc_basic@suspend-read-crc-pipe-*) on ILK/SNB/IVB, affecting CI.

This happened with v1 but not with v2 of the patch. Please check if somehow v1 ended up in your tree.

v2 passed CI fine, save for one warning not caused by the patch:
https://patchwork.freedesktop.org/series/4068/

Quoted 48 lines

For comparison, this was v1:
https://patchwork.freedesktop.org/patch/75840/

Best regards,

Lukas

Quoted 41 lines

Probably the same problem with the one in v2, but on older HW.

commit a7442b93cf32c1e1ddb721a26cd1f92302e2a222

Author: Lukas Wunner <[email protected]>
Date:   Wed Mar 9 12:52:53 2016 +0100

drm/i915: Fix races on fbdev

The ->lastclose callback invokes intel_fbdev_restore_mode() and has been witnessed to run before intel_fbdev_initial_config_async() has finished.

We might likewise receive hotplug events before we've had a chance to fully set up the fbdev.

Fix by waiting for the asynchronous thread to finish.

v2:
An async_synchronize_full() was also added to intel_fbdev_set_suspend()
in v1 which turned out to be entirely gratuitous. It caused a deadlock on suspend (discovered by CI, thanks to Damien Lespiau and Tomi Sarvela
for CI support) and was unnecessary since a device is never suspended until its ->probe callback (and all asynchronous tasks it scheduled) have finished. See dpm_prepare(), which calls wait_for_device_probe(), which calls async_synchronize_full().

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

Link:
http://patchwork.freedesktop.org/patch/msgid/20160309115147.67B2B6E0D
[email protected]>
Regards,
Gabriel

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#22 lore raw
In reply to Lukas Wunner
Quoted 8 lines · Lukas Wunner

Hi Gabriel,

Quoted 2 lines · Gabriel Feceoru

This commit causes a hang while running kms suspend tests (kms_pipe_crc_basic@suspend-read-crc-pipe-*) on ILK/SNB/IVB, affecting CI.

This happened with v1 but not with v2 of the patch. Please check if somehow v1 ended up in your tree.

It's v2.

Tomi already replied, meantime I also looked at the results.
The current regression is for ILK/SNB/IVB only (v1 seemed to affect more platforms). Unfortunately these machines were not available when v2 was tested, so this couldn't be detected.

Regards,
Gabriel.

Quoted 4 lines

v2 passed CI fine, save for one warning not caused by the patch:
https://patchwork.freedesktop.org/series/4068/

Quoted 47 lines

For comparison, this was v1:
https://patchwork.freedesktop.org/patch/75840/

Best regards,

Lukas

Quoted 40 lines

Probably the same problem with the one in v2, but on older HW.

commit a7442b93cf32c1e1ddb721a26cd1f92302e2a222

Author: Lukas Wunner <[email protected]>
Date:   Wed Mar 9 12:52:53 2016 +0100

drm/i915: Fix races on fbdev

The ->lastclose callback invokes intel_fbdev_restore_mode() and has been witnessed to run before intel_fbdev_initial_config_async() has finished.

We might likewise receive hotplug events before we've had a chance to fully set up the fbdev.

Fix by waiting for the asynchronous thread to finish.

v2:
An async_synchronize_full() was also added to intel_fbdev_set_suspend() in v1 which turned out to be entirely gratuitous. It caused a deadlock on suspend (discovered by CI, thanks to Damien Lespiau and Tomi Sarvela for CI support) and was unnecessary since a device is never suspended until its ->probe callback (and all asynchronous tasks it scheduled) have finished. See dpm_prepare(), which calls wait_for_device_probe(), which calls async_synchronize_full().

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

Regards,
Gabriel

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#7 lore raw

Re: [PATCH v3] drm/i915: Protect fbdev across slow or failed initialisation

In reply to Chris Wilson

This almost fixes the problem , but with this:

drivers/gpu/drm/i915/intel_fbdev.c +1 -1
index 5029f92..a6d3c58 100644
@@ -123,8 +123,9 @@ static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
                return NULL;
 
        info = dev_priv->fbdev->helper.fbdev;
-       if (info->screen_base == NULL)
+       if (info == NULL || info->screen_base == NULL)
                return NULL;
 
        if (info->state != FBINFO_STATE_RUNNING)
                return NULL;

I got initially a page fault during init, but with this fix, the suspend test passes.

Thank you, please consider this.

Regards,
Gabriel

Quoted 87 lines · Chris Wilson

If the initialisation fails, we may be left with a dangling pointer with an incomplete fbdev structure. Here we want to disable internal calls into fbdev. Similarly, the initialisation may be slow and we haven't yet enabled the fbdev (e.g. quick suspend or last-close before the async init completes).

v3: To create a typo introduced when retyping

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

---
  drivers/gpu/drm/i915/intel_fbdev.c | 48 ++++++++++++++++++++++++--------------
  1 file changed, 30 insertions(+), 18 deletions(-)
drivers/gpu/drm/i915/intel_fbdev.c +27 -14
index 153ea7a3fcf6..5029f927fe0d 100644
@@ -114,6 +114,24 @@ static struct fb_ops intelfb_ops = {
  	.fb_debug_leave = drm_fb_helper_debug_leave,
  };
 
+static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
+{
+	struct drm_i915_private *dev_priv = to_i915(dev);
+	struct fb_info *info;
+
+	if (dev_priv->fbdev == NULL)
+		return NULL;
+
+	info = dev_priv->fbdev->helper.fbdev;
+	if (info->screen_base == NULL)
+		return NULL;
+
+	if (info->state != FBINFO_STATE_RUNNING)
+		return NULL;
+
+	return dev_priv->fbdev;
+}
+
  static int intelfb_alloc(struct drm_fb_helper *helper,
  			 struct drm_fb_helper_surface_size *sizes)
  {
@@ -766,6 +784,8 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
  		return;
 
  	info = ifbdev->helper.fbdev;
+	if (info->screen_base == NULL)
+		return;
 
  	if (synchronous) {
  		/* Flush any pending work to turn the console on, and then
@@ -807,32 +827,24 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
 
  void intel_fbdev_output_poll_changed(struct drm_device *dev)
  {
-	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
+
+	if (ifbdev == NULL)
+		return;
 
-	async_synchronize_full();
-	if (dev_priv->fbdev)
-		drm_fb_helper_hotplug_event(&dev_priv->fbdev->helper);
+	drm_fb_helper_hotplug_event(&ifbdev->helper);
  }
 
  void intel_fbdev_restore_mode(struct drm_device *dev)
  {
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
 
-	async_synchronize_full();
-	if (!ifbdev)
+	if (ifbdev == NULL)
  		return;
 
-	fb_helper = &ifbdev->helper;
-
-	ret = drm_fb_helper_restore_fbdev_mode_unlocked(fb_helper);
-	if (ret) {

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous · line 850

Quoted 10 lines
drivers/gpu/drm/i915/intel_fbdev.c +3 -4
-		DRM_DEBUG("failed to restore crtc mode\n");
-	} else {
-		mutex_lock(&fb_helper->dev->struct_mutex);
+	if (drm_fb_helper_restore_fbdev_mode_unlocked(&ifbdev->helper) == 0) {
+		mutex_lock(&dev->struct_mutex);
  		intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);
-		mutex_unlock(&fb_helper->dev->struct_mutex);
+		mutex_unlock(&dev->struct_mutex);
  	}
  }

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#8 lore raw

PATCH v4 1/2 drm/i915: Protect fbdev across slow or failed initialisation

If the initialisation fails, we may be left with a dangling pointer with an incomplete fbdev structure. Here we want to disable internal calls into fbdev. Similarly, the initialisation may be slow and we haven't yet enabled the fbdev (e.g. quick suspend or last-close before the async init completes).

v3: To create a typo introduced when retyping
v4: Prevent info==NULL dereference in early boot

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

---
 drivers/gpu/drm/i915/intel_fbdev.c | 72 +++++++++++++++++++++++++-------------
 1 file changed, 48 insertions(+), 24 deletions(-)
drivers/gpu/drm/i915/intel_fbdev.c +48 -24
index 153ea7a3fcf6..5d4be71bdf22 100644
@@ -756,17 +756,47 @@ void intel_fbdev_fini(struct drm_device *dev)
 	dev_priv->fbdev = NULL;
 }
 
+static struct intel_fbdev *intel_fbdev_get(struct drm_device *dev)
+{
+	struct drm_i915_private *dev_priv = to_i915(dev);
+	struct fb_info *info;
+
+	if (dev_priv->fbdev == NULL)
+		return NULL;
+
+	info = dev_priv->fbdev->helper.fbdev;
+	if (info == NULL)
+		return NULL;
+
+	if (info->screen_base == NULL)
+		return NULL;
+
+	return dev_priv->fbdev;
+}
+
+static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
+{
+	struct intel_fbdev *ifbdev;
+
+	ifbdev = intel_fbdev_get(dev);
+	if (ifbdev == NULL)
+		return NULL;
+
+	if (ifbdev->helper.fbdev->state != FBINFO_STATE_RUNNING)
+		return NULL;
+
+	return ifbdev;
+}
+
 void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct fb_info *info;
+	struct intel_fbdev *ifbdev;
 
-	if (!ifbdev)
+	ifbdev = intel_fbdev_get(dev);
+	if (ifbdev == NULL)
 		return;
 
-	info = ifbdev->helper.fbdev;
-
 	if (synchronous) {
 		/* Flush any pending work to turn the console on, and then
 		 * wait to turn it off. It must be synchronous as we are
@@ -798,8 +828,10 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
 	 * been restored from swap. If the object is stolen however, it will be
 	 * full of whatever garbage was left in there.
 	 */
-	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen)
+	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen) {
+		struct fb_info *info = ifbdev->helper.fbdev;
 		memset_io(info->screen_base, 0, info->screen_size);
+	}
 
 	drm_fb_helper_set_suspend(&ifbdev->helper, state);
 	console_unlock();
@@ -807,32 +839,24 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
 
 void intel_fbdev_output_poll_changed(struct drm_device *dev)
 {
-	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
+
+	if (ifbdev == NULL)
+		return;
 
-	async_synchronize_full();
-	if (dev_priv->fbdev)
-		drm_fb_helper_hotplug_event(&dev_priv->fbdev->helper);
+	drm_fb_helper_hotplug_event(&ifbdev->helper);
 }
 
 void intel_fbdev_restore_mode(struct drm_device *dev)
 {
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
 
-	async_synchronize_full();
-	if (!ifbdev)
+	if (ifbdev == NULL)
 		return;
 
-	fb_helper = &ifbdev->helper;
-
-	ret = drm_fb_helper_restore_fbdev_mode_unlocked(fb_helper);
-	if (ret) {
-		DRM_DEBUG("failed to restore crtc mode\n");
-	} else {
-		mutex_lock(&fb_helper->dev->struct_mutex);
+	if (drm_fb_helper_restore_fbdev_mode_unlocked(&ifbdev->helper) == 0) {
+		mutex_lock(&dev->struct_mutex);
 		intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);
-		mutex_unlock(&fb_helper->dev->struct_mutex);
+		mutex_unlock(&dev->struct_mutex);
 	}
 }
Signature · 6 lines
2.8.0.rc3

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#9 lore raw

PATCH v4 2/2 drm/i915: Move fbdev_suspend_work to intel_fbdev

Since the suspend_work is entirely internal to intel_fbdev.c, move it from the top level drm_i915_private and into struct intel_fbdev. This requires splitting the internal interface for the suspend worker from the external interface used by the higher layers to initiate suspend.

---
 drivers/gpu/drm/i915/i915_drv.c    |   4 +-
 drivers/gpu/drm/i915/i915_drv.h    |   1 -
 drivers/gpu/drm/i915/intel_drv.h   |   5 +-
 drivers/gpu/drm/i915/intel_fbdev.c | 117 ++++++++++++++++++++-----------------
 4 files changed, 68 insertions(+), 59 deletions(-)
drivers/gpu/drm/i915/i915_drv.c +2 -2
index 020a31c5e2bb..100d0d92b1e6 100644
@@ -634,7 +634,7 @@ static int i915_drm_suspend(struct drm_device *dev)
 	intel_uncore_forcewake_reset(dev, false);
 	intel_opregion_fini(dev);
 
-	intel_fbdev_set_suspend(dev, FBINFO_STATE_SUSPENDED, true);
+	intel_fbdev_set_suspend(dev, FBINFO_STATE_SUSPENDED);
 
 	dev_priv->suspend_count++;
 
@@ -784,7 +784,7 @@ static int i915_drm_resume(struct drm_device *dev)
 
 	intel_opregion_init(dev);
 
-	intel_fbdev_set_suspend(dev, FBINFO_STATE_RUNNING, false);
+	intel_fbdev_set_suspend(dev, FBINFO_STATE_RUNNING);
 
 	mutex_lock(&dev_priv->modeset_restore_lock);
 	dev_priv->modeset_restore = MODESET_DONE;
drivers/gpu/drm/i915/i915_drv.h +0 -1
index 820c91f551ba..973a602c5077 100644
@@ -1880,7 +1880,6 @@ struct drm_i915_private {
 #ifdef CONFIG_DRM_FBDEV_EMULATION
 	/* list of fbdev register on this device */
 	struct intel_fbdev *fbdev;
-	struct work_struct fbdev_suspend_work;
 #endif
 
 	struct drm_property *broadcast_rgb_property;
drivers/gpu/drm/i915/intel_drv.h +3 -2
index 6ac46d921cde..00b6c60c1cb8 100644
@@ -158,6 +158,7 @@ struct intel_framebuffer {
 struct intel_fbdev {
 	struct drm_fb_helper helper;
 	struct intel_framebuffer *fb;
+	struct work_struct suspend_work;
 	int preferred_bpp;
 };
 
@@ -1335,7 +1336,7 @@ void intel_dvo_init(struct drm_device *dev);
 extern int intel_fbdev_init(struct drm_device *dev);
 extern void intel_fbdev_initial_config_async(struct drm_device *dev);
 extern void intel_fbdev_fini(struct drm_device *dev);
-extern void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous);
+extern void intel_fbdev_set_suspend(struct drm_device *dev, int state);
 extern void intel_fbdev_output_poll_changed(struct drm_device *dev);
 extern void intel_fbdev_restore_mode(struct drm_device *dev);
 #else
@@ -1352,7 +1353,7 @@ static inline void intel_fbdev_fini(struct drm_device *dev)
 {
 }
 
-static inline void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous)
+static inline void intel_fbdev_set_suspend(struct drm_device *dev, int state)
 {
 }
 
drivers/gpu/drm/i915/intel_fbdev.c +63 -54

163 lines not on this page · raw message

Signature · 6 lines
2.8.0.rc3

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#24 lore raw

✗ Fi.CI.BAT: failure for drm/i915: Protect fbdev across slow or failed initialisation (rev2)

In reply to Gabriel Feceoru
== Series Details ==

== Series Details ==

Series: drm/i915: Protect fbdev across slow or failed initialisation (rev2)
URL : https://patchwork.freedesktop.org/series/5065/
State : failure

== Summary ==

Series 5065v2 drm/i915: Protect fbdev across slow or failed initialisation
http://patchwork.freedesktop.org/api/1.0/series/5065/revisions/2/mbox/

Test drv_module_reload_basic:

                pass       -> INCOMPLETE (bsw-nuc-2)
                pass       -> INCOMPLETE (skl-nuci5)
                pass       -> INCOMPLETE (byt-nuc)
Test gem_exec_suspend:
        Subgroup basic-s3:
                pass       -> DMESG-WARN (bsw-nuc-2)
Test kms_pipe_crc_basic:
        Subgroup suspend-read-crc-pipe-c:
                dmesg-warn -> PASS       (bsw-nuc-2)
Test pm_rpm:
        Subgroup basic-rte:
                pass       -> DMESG-WARN (bsw-nuc-2)
bsw-nuc-2        total:154  pass:122  dwarn:2   dfail:0   fail:0   skip:29 
byt-nuc          total:57   pass:46   dwarn:0   dfail:0   fail:0   skip:10 
hsw-brixbox      total:196  pass:174  dwarn:0   dfail:0   fail:0   skip:22 
skl-nuci5        total:25   pass:22   dwarn:0   dfail:0   fail:0   skip:2  

Results at /archive/results/CI_IGT_test/Patchwork_1761/

03c0f854e93263563f559d2bc8e47fb51adae697 drm-intel-nightly: 2016y-03m-31d-10h-50m-15s UTC integration manifest ab3a2f88b9f8b6a3109c325d42770c3c5fd714a4 drm/i915: Protect fbdev across slow or failed initialisation

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#23 lore raw
In reply to Gabriel Feceoru

Hi Gabriel,

Quoted 10 lines · Gabriel Feceoru
Quoted 3 lines · Lukas Wunner
Quoted 2 lines · Gabriel Feceoru

This commit causes a hang while running kms suspend tests (kms_pipe_crc_basic@suspend-read-crc-pipe-*) on ILK/SNB/IVB, affecting CI.

Tomi already replied, meantime I also looked at the results. The current regression is for ILK/SNB/IVB only (v1 seemed to affect more platforms).
Unfortunately these machines were not available when v2 was tested, so this couldn't be detected.

I dev on an IVB machine and cannot reproduce this. Suspend works fine.

All the patch does is call async_synchronize_full()
(1) when a hotplug event arrives or
(2) when the last DRM client closes the connection.
Either of these two things seems to be happening on your test machines when running the suspend test.

The PM core suspends and resumes individual devices asynchronously and calls async_synchronize_full() in a couple of places. If a device's PM callbacks also call async_synchronize_full(), the machine deadlocks.

It is unnecessary that we call async_synchronize_full(), we only need to synchronize up to a specific cookie (which represents initialization of the fbdev). So I've just posted a patch to replace the calls to async_synchronize_full() with async_synchronize_cookie(). This should make things less fragile and hopefully also solve the hangs you're seeing.

Best regards,

Lukas

Quoted 38 lines
Quoted 38 lines
Quoted 38 lines

Probably the same problem with the one in v2, but on older HW.

commit a7442b93cf32c1e1ddb721a26cd1f92302e2a222

Author: Lukas Wunner <[email protected]>
Date:   Wed Mar 9 12:52:53 2016 +0100

drm/i915: Fix races on fbdev

The ->lastclose callback invokes intel_fbdev_restore_mode() and has been witnessed to run before intel_fbdev_initial_config_async() has finished.

We might likewise receive hotplug events before we've had a chance to fully set up the fbdev.

Fix by waiting for the asynchronous thread to finish.

v2:
An async_synchronize_full() was also added to intel_fbdev_set_suspend() in v1 which turned out to be entirely gratuitous. It caused a deadlock on suspend (discovered by CI, thanks to Damien Lespiau and Tomi Sarvela for CI support) and was unnecessary since a device is never suspended until its ->probe callback (and all asynchronous tasks it scheduled) have finished. See dpm_prepare(), which calls wait_for_device_probe(), which calls async_synchronize_full().

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

Quoted 8 lines
Quoted 8 lines
Quoted 2 lines

Regards,
Gabriel

v2 passed CI fine, save for one warning not caused by the patch:
https://patchwork.freedesktop.org/series/4068/

For comparison, this was v1:
https://patchwork.freedesktop.org/patch/75840/

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

maintainer Reviewed-by on v4 2/2
#10 lore raw

Re: [PATCH v4 2/2] drm/i915: Move fbdev_suspend_work to intel_fbdev

In reply to Chris Wilson
Quoted 140 lines · Chris Wilson

Since the suspend_work is entirely internal to intel_fbdev.c, move it from the top level drm_i915_private and into struct intel_fbdev. This requires splitting the internal interface for the suspend worker from the external interface used by the higher layers to initiate suspend.

---
 drivers/gpu/drm/i915/i915_drv.c    |   4 +- drivers/gpu/drm/i915/i915_drv.h    |   1 - drivers/gpu/drm/i915/intel_drv.h   |   5 +- drivers/gpu/drm/i915/intel_fbdev.c | 117 ++++++++++++++++++++-----------------

 4 files changed, 68 insertions(+), 59 deletions(-)
drivers/gpu/drm/i915/i915_drv.c +0 -0
index 020a31c5e2bb..100d0d92b1e6 100644
@@ -634,7 +634,7 @@ static int i915_drm_suspend(struct drm_device *dev)
  	intel_uncore_forcewake_reset(dev, false);
  	intel_opregion_fini(dev);

drivers/gpu/drm/i915/i915_drv.c +1 -1
-	intel_fbdev_set_suspend(dev, FBINFO_STATE_SUSPENDED, true);
+	intel_fbdev_set_suspend(dev, FBINFO_STATE_SUSPENDED);


  dev_priv->suspend_count++;

drivers/gpu/drm/i915/i915_drv.c +0 -0
@@ -784,7 +784,7 @@ static int i915_drm_resume(struct drm_device *dev)


  intel_opregion_init(dev);

drivers/gpu/drm/i915/i915_drv.c +1 -1
-	intel_fbdev_set_suspend(dev, FBINFO_STATE_RUNNING, false);
+	intel_fbdev_set_suspend(dev, FBINFO_STATE_RUNNING);

mutex_lock(&dev_priv->modeset_restore_lock); dev_priv->modeset_restore = MODESET_DONE;

drivers/gpu/drm/i915/i915_drv.h +0 -1
index 820c91f551ba..973a602c5077 100644
@@ -1880,7 +1880,6 @@ struct drm_i915_private {
  #ifdef CONFIG_DRM_FBDEV_EMULATION
  	/* list of fbdev register on this device */
  	struct intel_fbdev *fbdev;
-	struct work_struct fbdev_suspend_work;
  #endif


  struct drm_property *broadcast_rgb_property;

drivers/gpu/drm/i915/intel_drv.h +1 -0
index 6ac46d921cde..00b6c60c1cb8 100644
@@ -158,6 +158,7 @@ struct intel_framebuffer {
  struct intel_fbdev {
  	struct drm_fb_helper helper;
  	struct intel_framebuffer *fb;
+	struct work_struct suspend_work;
  	int preferred_bpp;
  };

drivers/gpu/drm/i915/intel_drv.h +1 -1
@@ -1335,7 +1336,7 @@ void intel_dvo_init(struct drm_device *dev);
  extern int intel_fbdev_init(struct drm_device *dev);
  extern void intel_fbdev_initial_config_async(struct drm_device *dev);
  extern void intel_fbdev_fini(struct drm_device *dev);
-extern void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous);
+extern void intel_fbdev_set_suspend(struct drm_device *dev, int state);
  extern void intel_fbdev_output_poll_changed(struct drm_device *dev);
  extern void intel_fbdev_restore_mode(struct drm_device *dev);
  #else
@@ -1352,7 +1353,7 @@ static inline void intel_fbdev_fini(struct drm_device *dev)
  {
  }

drivers/gpu/drm/i915/intel_drv.h +1 -1
-static inline void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous)
+static inline void intel_fbdev_set_suspend(struct drm_device *dev, int state)

 {
 }

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
index 5d4be71bdf22..66bb79613660 100644
@@ -534,8 +534,7 @@ static const struct drm_fb_helper_funcs intel_fb_helper_funcs = {
  	.fb_probe = intelfb_create,
  };

drivers/gpu/drm/i915/intel_fbdev.c +1 -2
-static void intel_fbdev_destroy(struct drm_device *dev,
-				struct intel_fbdev *ifbdev)
+static void intel_fbdev_destroy(struct intel_fbdev *ifbdev)

 {
  /* We rely on the object-free to release the VMA pinning for
   * the info->screen_base mmaping. Leaking the VMA is simpler than

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
@@ -681,13 +680,56 @@ out:
  	return false;
  }

drivers/gpu/drm/i915/intel_fbdev.c +43 -0
+static void __intel_fbdev_set_suspend(struct intel_fbdev *ifbdev,
+				      int state, bool synchronous)
+{
+	if (synchronous) {
+		/* Flush any pending work to turn the console on, and then
+		 * wait to turn it off. It must be synchronous as we are
+		 * about to suspend or unload the driver.
+		 *
+		 * Note that from within the work-handler, we cannot flush
+		 * ourselves, so only flush outstanding work upon suspend!
+		 */
+		if (state != FBINFO_STATE_RUNNING)
+			flush_work(&ifbdev->suspend_work);
+		console_lock();
+	} else {
+		/*
+		 * The console lock can be pretty contented on resume due
+		 * to all the printk activity.  Try to keep it out of the hot
+		 * path of resume if possible.
+		 */
+		WARN_ON(state != FBINFO_STATE_RUNNING);
+		if (!console_trylock()) {
+			/* Don't block our own workqueue as this can
+			 * be run in parallel with other i915.ko tasks.
+			 */
+			schedule_work(&ifbdev->suspend_work);
+			return;
+		}
+	}
+
+	/* On resume from hibernation: If the object is shmemfs backed, it has
+	 * been restored from swap. If the object is stolen however, it will be
+	 * full of whatever garbage was left in there.
+	 */
+	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen) {
+		struct fb_info *info = ifbdev->helper.fbdev;
+		memset_io(info->screen_base, 0, info->screen_size);
+	}
+
+	drm_fb_helper_set_suspend(&ifbdev->helper, state);
+	console_unlock();
+}
+

 static void intel_fbdev_suspend_worker(struct work_struct *work) {

drivers/gpu/drm/i915/intel_fbdev.c out: · line 730

Quoted 8 lines
drivers/gpu/drm/i915/intel_fbdev.c +3 -5
-	intel_fbdev_set_suspend(container_of(work,
-					     struct drm_i915_private,
-					     fbdev_suspend_work)->dev,
-				FBINFO_STATE_RUNNING,
-				true);
+	__intel_fbdev_set_suspend(container_of(work,
+					       struct intel_fbdev,
+					       suspend_work),

Have the container_of on separate line at least, maybe even with a macro.

Quoted 20 lines
drivers/gpu/drm/i915/intel_fbdev.c +2 -0
+				  FBINFO_STATE_RUNNING,
+				  true);

 }

 int intel_fbdev_init(struct drm_device *dev)

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
@@ -716,9 +758,9 @@ int intel_fbdev_init(struct drm_device *dev)
  	}


  ifbdev->helper.atomic = true;

drivers/gpu/drm/i915/intel_fbdev.c +1 -0
+	INIT_WORK(&ifbdev->suspend_work, intel_fbdev_suspend_worker);


  dev_priv->fbdev = ifbdev;

drivers/gpu/drm/i915/intel_fbdev.c +0 -1
-	INIT_WORK(&dev_priv->fbdev_suspend_work, intel_fbdev_suspend_worker);

drm_fb_helper_single_add_all_connectors(&ifbdev->helper);

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
@@ -743,17 +785,21 @@ void intel_fbdev_initial_config_async(struct drm_device *dev)


 void intel_fbdev_fini(struct drm_device *dev)
 {

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_initial_config_async(struct drm_device *dev) · line 791

Quoted 6 lines
drivers/gpu/drm/i915/intel_fbdev.c +4 -2
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	if (!dev_priv->fbdev)
+	struct drm_i915_private *dev_priv = to_i915(dev);
+	struct intel_fbdev *ifbdev;
+
+	ifbdev = dev_priv->fbdev;

Straight assignment.

Quoted 62 lines
drivers/gpu/drm/i915/intel_fbdev.c +1 -0
+	if (ifbdev == NULL)

  return;

drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-	flush_work(&dev_priv->fbdev_suspend_work);
+	dev_priv->fbdev = NULL;


  if (!current_is_async())
  async_synchronize_full();

drivers/gpu/drm/i915/intel_fbdev.c +4 -3
-	intel_fbdev_destroy(dev, dev_priv->fbdev);
-	kfree(dev_priv->fbdev);
-	dev_priv->fbdev = NULL;
+	flush_work(&ifbdev->suspend_work);
+
+	intel_fbdev_destroy(ifbdev);
+	kfree(ifbdev);

 }

 static struct intel_fbdev *intel_fbdev_get(struct drm_device *dev)

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
@@ -788,53 +834,16 @@ static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
  	return ifbdev;
  }

drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous)
+void intel_fbdev_set_suspend(struct drm_device *dev, int state)

 {

drivers/gpu/drm/i915/intel_fbdev.c +0 -1
-	struct drm_i915_private *dev_priv = dev->dev_private;

  struct intel_fbdev *ifbdev;

  ifbdev = intel_fbdev_get(dev);
  if (ifbdev == NULL)
  return;

drivers/gpu/drm/i915/intel_fbdev.c +0 -30
-	if (synchronous) {
-		/* Flush any pending work to turn the console on, and then
-		 * wait to turn it off. It must be synchronous as we are
-		 * about to suspend or unload the driver.
-		 *
-		 * Note that from within the work-handler, we cannot flush
-		 * ourselves, so only flush outstanding work upon suspend!
-		 */
-		if (state != FBINFO_STATE_RUNNING)
-			flush_work(&dev_priv->fbdev_suspend_work);
-		console_lock();
-	} else {
-		/*
-		 * The console lock can be pretty contented on resume due
-		 * to all the printk activity.  Try to keep it out of the hot
-		 * path of resume if possible.
-		 */
-		WARN_ON(state != FBINFO_STATE_RUNNING);
-		if (!console_trylock()) {
-			/* Don't block our own workqueue as this can
-			 * be run in parallel with other i915.ko tasks.
-			 */
-			schedule_work(&dev_priv->fbdev_suspend_work);
-			return;
-		}
-	}
-
-	/* On resume from hibernation: If the object is shmemfs backed, it has
-	 * been restored from swap. If the object is stolen however, it will be
-	 * full of whatever garbage was left in there.

drivers/gpu/drm/i915/intel_fbdev.c static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev) · line 846

Quoted 10 lines
drivers/gpu/drm/i915/intel_fbdev.c +2 -8
-	 */
-	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen) {
-		struct fb_info *info = ifbdev->helper.fbdev;
-		memset_io(info->screen_base, 0, info->screen_size);
-	}
-
-	drm_fb_helper_set_suspend(&ifbdev->helper, state);
-	console_unlock();
+	__intel_fbdev_set_suspend(ifbdev, state,
+				  state != FBINFO_STATE_RUNNING);

Could have changed this function name to make it easier to spot what you changed or not. Code motion as separate patches.

With above addressed,

Quoted 3 lines

 }

 void intel_fbdev_output_poll_changed(struct drm_device *dev)

Signature · 7 lines
Joonas Lahtinen
Open Source Technology Center
Intel Corporation
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#11 lore raw

Re: [PATCH v4 2/2] drm/i915: Move fbdev_suspend_work to intel_fbdev

Quoted 14 lines · Joonas Lahtinen

drivers/gpu/drm/i915/intel_fbdev.c out: · line 730

Quoted 10 lines · Chris Wilson

 static void intel_fbdev_suspend_worker(struct work_struct *work) {

drivers/gpu/drm/i915/intel_fbdev.c +3 -5
-	intel_fbdev_set_suspend(container_of(work,
-					     struct drm_i915_private,
-					     fbdev_suspend_work)->dev,
-				FBINFO_STATE_RUNNING,
-				true);
+	__intel_fbdev_set_suspend(container_of(work,
+					       struct intel_fbdev,
+					       suspend_work),

Have the container_of on separate line at least, maybe even with a macro.

Sure.

Quoted 29 lines
Quoted 20 lines
drivers/gpu/drm/i915/intel_fbdev.c +2 -0
+				  FBINFO_STATE_RUNNING,
+				  true);

 }

 int intel_fbdev_init(struct drm_device *dev)

(quoted diff) +0 -0
@@ -716,9 +758,9 @@ int intel_fbdev_init(struct drm_device *dev)
  	}


  ifbdev->helper.atomic = true;

drivers/gpu/drm/i915/intel_fbdev.c +1 -0
+	INIT_WORK(&ifbdev->suspend_work, intel_fbdev_suspend_worker);


  dev_priv->fbdev = ifbdev;

drivers/gpu/drm/i915/intel_fbdev.c +0 -1
-	INIT_WORK(&dev_priv->fbdev_suspend_work, intel_fbdev_suspend_worker);

drm_fb_helper_single_add_all_connectors(&ifbdev->helper);

(quoted diff) +0 -0
@@ -743,17 +785,21 @@ void intel_fbdev_initial_config_async(struct drm_device *dev)


 void intel_fbdev_fini(struct drm_device *dev)
 {

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_initial_config_async(struct drm_device *dev) · line 791

Quoted 6 lines
drivers/gpu/drm/i915/intel_fbdev.c +4 -2
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	if (!dev_priv->fbdev)
+	struct drm_i915_private *dev_priv = to_i915(dev);
+	struct intel_fbdev *ifbdev;
+
+	ifbdev = dev_priv->fbdev;

Straight assignment.

Quoted 1 line

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_initial_config_async(struct drm_device *dev) · line 792

Quoted 1 line
drivers/gpu/drm/i915/intel_fbdev.c +1 -0
+	if (ifbdev == NULL)

I was trying to group the local assignment with the if, and I like to leave whitespace after the locals. By placing these two lines together this function looks like the majority of the other functions.

Quoted 66 lines
Quoted 66 lines

  return;

drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-	flush_work(&dev_priv->fbdev_suspend_work);
+	dev_priv->fbdev = NULL;


  if (!current_is_async())
  async_synchronize_full();

drivers/gpu/drm/i915/intel_fbdev.c +4 -3
-	intel_fbdev_destroy(dev, dev_priv->fbdev);
-	kfree(dev_priv->fbdev);
-	dev_priv->fbdev = NULL;
+	flush_work(&ifbdev->suspend_work);
+
+	intel_fbdev_destroy(ifbdev);
+	kfree(ifbdev);

 }

 static struct intel_fbdev *intel_fbdev_get(struct drm_device *dev)

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
@@ -788,53 +834,16 @@ static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
  	return ifbdev;
  }

drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous)
+void intel_fbdev_set_suspend(struct drm_device *dev, int state)

 {

(quoted diff) +0 -1
-	struct drm_i915_private *dev_priv = dev->dev_private;

  struct intel_fbdev *ifbdev;

  ifbdev = intel_fbdev_get(dev);
  if (ifbdev == NULL)
  return;

drivers/gpu/drm/i915/intel_fbdev.c +0 -35
-	if (synchronous) {
-		/* Flush any pending work to turn the console on, and then
-		 * wait to turn it off. It must be synchronous as we are
-		 * about to suspend or unload the driver.
-		 *
-		 * Note that from within the work-handler, we cannot flush
-		 * ourselves, so only flush outstanding work upon suspend!
-		 */
-		if (state != FBINFO_STATE_RUNNING)
-			flush_work(&dev_priv->fbdev_suspend_work);
-		console_lock();
-	} else {
-		/*
-		 * The console lock can be pretty contented on resume due
-		 * to all the printk activity.  Try to keep it out of the hot
-		 * path of resume if possible.
-		 */
-		WARN_ON(state != FBINFO_STATE_RUNNING);
-		if (!console_trylock()) {
-			/* Don't block our own workqueue as this can
-			 * be run in parallel with other i915.ko tasks.
-			 */
-			schedule_work(&dev_priv->fbdev_suspend_work);
-			return;
-		}
-	}
-
-	/* On resume from hibernation: If the object is shmemfs backed, it has
-	 * been restored from swap. If the object is stolen however, it will be
-	 * full of whatever garbage was left in there.
-	 */
-	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen) {
-		struct fb_info *info = ifbdev->helper.fbdev;
-		memset_io(info->screen_base, 0, info->screen_size);
-	}
Quoted 8 lines

drivers/gpu/drm/i915/intel_fbdev.c static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev) · line 846

Quoted 5 lines
drivers/gpu/drm/i915/intel_fbdev.c +2 -3
-
-	drm_fb_helper_set_suspend(&ifbdev->helper, state);
-	console_unlock();
+	__intel_fbdev_set_suspend(ifbdev, state,
+				  state != FBINFO_STATE_RUNNING);

Could have changed this function name to make it easier to spot what you changed or not. Code motion as separate patches.

This was a separate patch! The purpose of this patch was to move code/data about :)
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#12 lore raw

Re: [PATCH v4 2/2] drm/i915: Move fbdev_suspend_work to intel_fbdev

Quoted 129 lines · Chris Wilson
Quoted 15 lines · Joonas Lahtinen

drivers/gpu/drm/i915/intel_fbdev.c out: · line 730

Quoted 11 lines · Chris Wilson

 static void intel_fbdev_suspend_worker(struct work_struct *work) {

drivers/gpu/drm/i915/intel_fbdev.c +3 -5
-	intel_fbdev_set_suspend(container_of(work,
-					     struct drm_i915_private,
-					     fbdev_suspend_work)->dev,
-				FBINFO_STATE_RUNNING,
-				true);
+	__intel_fbdev_set_suspend(container_of(work,
+					       struct intel_fbdev,
+					       suspend_work),

Have the container_of on separate line at least, maybe even with a macro.

Sure.

Quoted 31 lines
Quoted 22 lines
drivers/gpu/drm/i915/intel_fbdev.c +2 -0
+				  FBINFO_STATE_RUNNING,
+				  true);

 }

 int intel_fbdev_init(struct drm_device *dev)

(quoted diff) +0 -0
@@ -716,9 +758,9 @@ int intel_fbdev_init(struct drm_device *dev)
  	}


  ifbdev->helper.atomic = true;

drivers/gpu/drm/i915/intel_fbdev.c +1 -0
+	INIT_WORK(&ifbdev->suspend_work, intel_fbdev_suspend_worker);


  dev_priv->fbdev = ifbdev;

drivers/gpu/drm/i915/intel_fbdev.c +0 -1
-	INIT_WORK(&dev_priv->fbdev_suspend_work, intel_fbdev_suspend_worker);

drm_fb_helper_single_add_all_connectors(&ifbdev->helper);

(quoted diff) +0 -0
@@ -743,17 +785,21 @@ void intel_fbdev_initial_config_async(struct drm_device *dev)


 void intel_fbdev_fini(struct drm_device *dev)
 {

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_initial_config_async(struct drm_device *dev) · line 791

Quoted 6 lines
drivers/gpu/drm/i915/intel_fbdev.c +4 -2
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	if (!dev_priv->fbdev)
+	struct drm_i915_private *dev_priv = to_i915(dev);
+	struct intel_fbdev *ifbdev;
+
+	ifbdev = dev_priv->fbdev;

Straight assignment.

Quoted 1 line
Quoted 1 line

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_initial_config_async(struct drm_device *dev) · line 792

Quoted 1 line
drivers/gpu/drm/i915/intel_fbdev.c +1 -0
+	if (ifbdev == NULL)

I was trying to group the local assignment with the if, and I like to leave whitespace after the locals. By placing these two lines together this function looks like the majority of the other functions.

Quoted 67 lines
Quoted 67 lines

  return;

drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-	flush_work(&dev_priv->fbdev_suspend_work);
+	dev_priv->fbdev = NULL;


  if (!current_is_async())
  async_synchronize_full();

drivers/gpu/drm/i915/intel_fbdev.c +4 -3
-	intel_fbdev_destroy(dev, dev_priv->fbdev);
-	kfree(dev_priv->fbdev);
-	dev_priv->fbdev = NULL;
+	flush_work(&ifbdev->suspend_work);
+
+	intel_fbdev_destroy(ifbdev);
+	kfree(ifbdev);

 }

 static struct intel_fbdev *intel_fbdev_get(struct drm_device *dev)

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
@@ -788,53 +834,16 @@ static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
  	return ifbdev;
  }

drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous)
+void intel_fbdev_set_suspend(struct drm_device *dev, int state)

 {

(quoted diff) +0 -1
-	struct drm_i915_private *dev_priv = dev->dev_private;

  struct intel_fbdev *ifbdev;

  ifbdev = intel_fbdev_get(dev);
  if (ifbdev == NULL)
  return;

drivers/gpu/drm/i915/intel_fbdev.c +0 -34
-	if (synchronous) {
-		/* Flush any pending work to turn the console on, and then
-		 * wait to turn it off. It must be synchronous as we are
-		 * about to suspend or unload the driver.
-		 *
-		 * Note that from within the work-handler, we cannot flush
-		 * ourselves, so only flush outstanding work upon suspend!
-		 */
-		if (state != FBINFO_STATE_RUNNING)
-			flush_work(&dev_priv->fbdev_suspend_work);
-		console_lock();
-	} else {
-		/*
-		 * The console lock can be pretty contented on resume due
-		 * to all the printk activity.  Try to keep it out of the hot
-		 * path of resume if possible.
-		 */
-		WARN_ON(state != FBINFO_STATE_RUNNING);
-		if (!console_trylock()) {
-			/* Don't block our own workqueue as this can
-			 * be run in parallel with other i915.ko tasks.
-			 */
-			schedule_work(&dev_priv->fbdev_suspend_work);
-			return;
-		}
-	}
-
-	/* On resume from hibernation: If the object is shmemfs backed, it has
-	 * been restored from swap. If the object is stolen however, it will be
-	 * full of whatever garbage was left in there.
-	 */
-	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen) {
-		struct fb_info *info = ifbdev->helper.fbdev;
-		memset_io(info->screen_base, 0, info->screen_size);
Quoted 8 lines

drivers/gpu/drm/i915/intel_fbdev.c static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev) · line 846

Quoted 6 lines
drivers/gpu/drm/i915/intel_fbdev.c +2 -4
-	}
-
-	drm_fb_helper_set_suspend(&ifbdev->helper, state);
-	console_unlock();
+	__intel_fbdev_set_suspend(ifbdev, state,
+				  state != FBINFO_STATE_RUNNING);

Could have changed this function name to make it easier to spot what you changed or not. Code motion as separate patches.

Quoted 2 lines

This was a separate patch! The purpose of this patch was to move code/data about :)

It also changed order of some stuff, so I got suspicious ;) But looks like the function stayed the same, so OK for me.

Quoted 2 lines

-Chris

Signature · 7 lines
Joonas Lahtinen
Open Source Technology Center
Intel Corporation
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer Reviewed-by on v4 1/2
#13 lore raw

Re: [PATCH v4 1/2] drm/i915: Protect fbdev across slow or failed initialisation

In reply to Chris Wilson
Quoted 30 lines · Chris Wilson

If the initialisation fails, we may be left with a dangling pointer with an incomplete fbdev structure. Here we want to disable internal calls into fbdev. Similarly, the initialisation may be slow and we haven't yet enabled the fbdev (e.g. quick suspend or last-close before the async init completes).

v3: To create a typo introduced when retyping
v4: Prevent info==NULL dereference in early boot

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93580

---
 drivers/gpu/drm/i915/intel_fbdev.c | 72 +++++++++++++++++++++++++-------------

 1 file changed, 48 insertions(+), 24 deletions(-)
drivers/gpu/drm/i915/intel_fbdev.c +0 -0
index 153ea7a3fcf6..5d4be71bdf22 100644
@@ -756,17 +756,47 @@ void intel_fbdev_fini(struct drm_device *dev)
  	dev_priv->fbdev = NULL;
  }

drivers/gpu/drm/i915/intel_fbdev.c +5 -0
+static struct intel_fbdev *intel_fbdev_get(struct drm_device *dev)
+{
+	struct drm_i915_private *dev_priv = to_i915(dev);
+	struct fb_info *info;
+

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_fini(struct drm_device *dev) · line 773

Quoted 10 lines
drivers/gpu/drm/i915/intel_fbdev.c +10 -0
+	if (dev_priv->fbdev == NULL)
+		return NULL;
+
+	info = dev_priv->fbdev->helper.fbdev;
+	if (info == NULL)
+		return NULL;
+
+	if (info->screen_base == NULL)
+		return NULL;
+

This is rather verbose to my liking, I'd rather be dropping those '== NULL' and convert to !. But either way to me.

Quoted 70 lines
drivers/gpu/drm/i915/intel_fbdev.c +17 -0
+	return dev_priv->fbdev;
+}
+
+static struct intel_fbdev *intel_fbdev_get_if_active(struct drm_device *dev)
+{
+	struct intel_fbdev *ifbdev;
+
+	ifbdev = intel_fbdev_get(dev);
+	if (ifbdev == NULL)
+		return NULL;
+
+	if (ifbdev->helper.fbdev->state != FBINFO_STATE_RUNNING)
+		return NULL;
+
+	return ifbdev;
+}
+

 void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous) {
  struct drm_i915_private *dev_priv = dev->dev_private;

drivers/gpu/drm/i915/intel_fbdev.c +1 -2
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct fb_info *info;
+	struct intel_fbdev *ifbdev;

drivers/gpu/drm/i915/intel_fbdev.c +2 -1
-	if (!ifbdev)
+	ifbdev = intel_fbdev_get(dev);
+	if (ifbdev == NULL)

  return;

drivers/gpu/drm/i915/intel_fbdev.c +0 -2
-	info = ifbdev->helper.fbdev;
-

  if (synchronous) {
  /* Flush any pending work to turn the console on, and then
   * wait to turn it off. It must be synchronous as we are

drivers/gpu/drm/i915/intel_fbdev.c +3 -1
@@ -798,8 +828,10 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous
  	 * been restored from swap. If the object is stolen however, it will be
  	 * full of whatever garbage was left in there.
  	 */
-	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen)
+	if (state == FBINFO_STATE_RUNNING && ifbdev->fb->obj->stolen) {
+		struct fb_info *info = ifbdev->helper.fbdev;
  		memset_io(info->screen_base, 0, info->screen_size);
+	}


  drm_fb_helper_set_suspend(&ifbdev->helper, state); console_unlock();

drivers/gpu/drm/i915/intel_fbdev.c +0 -0
@@ -807,32 +839,24 @@ void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous


 void intel_fbdev_output_poll_changed(struct drm_device *dev) {

drivers/gpu/drm/i915/intel_fbdev.c +4 -1
-	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);
+
+	if (ifbdev == NULL)
+		return;

drivers/gpu/drm/i915/intel_fbdev.c +1 -3
-	async_synchronize_full();
-	if (dev_priv->fbdev)
-		drm_fb_helper_hotplug_event(&dev_priv->fbdev->helper);
+	drm_fb_helper_hotplug_event(&ifbdev->helper);

 }

 void intel_fbdev_restore_mode(struct drm_device *dev) {

drivers/gpu/drm/i915/intel_fbdev.c +1 -4
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);

drivers/gpu/drm/i915/intel_fbdev.c

Quoted 1 line
drivers/gpu/drm/i915/intel_fbdev.c +0 -1
-	async_synchronize_full();

What's with the async_synchronize_full() begin removed completely?

drivers/gpu/drm/i915/intel_fbdev.c

Quoted 2 lines
drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-	if (!ifbdev)
+	if (ifbdev == NULL)

Argh.

Quoted 12 lines

  return;

drivers/gpu/drm/i915/intel_fbdev.c +2 -7
-	fb_helper = &ifbdev->helper;
-
-	ret = drm_fb_helper_restore_fbdev_mode_unlocked(fb_helper);
-	if (ret) {
-		DRM_DEBUG("failed to restore crtc mode\n");
-	} else {
-		mutex_lock(&fb_helper->dev->struct_mutex);
+	if (drm_fb_helper_restore_fbdev_mode_unlocked(&ifbdev->helper) == 0) {
+		mutex_lock(&dev->struct_mutex);

  intel_fb_obj_invalidate(ifbdev->fb->obj, ORIGIN_GTT);

drivers/gpu/drm/i915/intel_fbdev.c void intel_fbdev_set_suspend(struct drm_device *dev, int state, bool synchronous · line 860

Quoted 2 lines
drivers/gpu/drm/i915/intel_fbdev.c +1 -1
-		mutex_unlock(&fb_helper->dev->struct_mutex);
+		mutex_unlock(&dev->struct_mutex);

Above addressed,

Quoted 2 lines

  }
 }

Signature · 7 lines
Joonas Lahtinen
Open Source Technology Center
Intel Corporation
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#14 lore raw

Re: [PATCH v4 1/2] drm/i915: Protect fbdev across slow or failed initialisation

Quoted 12 lines · Joonas Lahtinen

Review

Quoted 9 lines · Chris Wilson

 void intel_fbdev_restore_mode(struct drm_device *dev) {

drivers/gpu/drm/i915/intel_fbdev.c +1 -4
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);

(quoted diff) +0 -1
-	async_synchronize_full();

What's with the async_synchronize_full() begin removed completely?

Because it's not just wrong, but completely broken imo.

During suspend, we want to freeze the async task not flush. Then here and during resume we skip the restoration of the unregistered fbdev, and then when the task is woken it can complete the registration and present the vanilla ifbdev.

During hibernation, we really just want to cancel the task and start from scratch on resume.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#15 lore raw

Re: [PATCH v4 1/2] drm/i915: Protect fbdev across slow or failed initialisation

Quoted 16 lines · Chris Wilson
Quoted 13 lines · Joonas Lahtinen

Review

Quoted 10 lines · Chris Wilson

 void intel_fbdev_restore_mode(struct drm_device *dev) {

drivers/gpu/drm/i915/intel_fbdev.c +1 -4
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);

(quoted diff) +0 -1
-	async_synchronize_full();

What's with the async_synchronize_full() begin removed completely?

Because it's not just wrong, but completely broken imo.

Quoted 7 lines

During suspend, we want to freeze the async task not flush. Then here and during resume we skip the restoration of the unregistered fbdev, and then when the task is woken it can complete the registration and present the vanilla ifbdev.

During hibernation, we really just want to cancel the task and start from scratch on resume.

Maybe Ack from Lukas with those? As this is effectively a revert of :

	commit a7442b93cf32c1e1ddb721a26cd1f92302e2a222
	Author: Lukas Wunner <[email protected]>
	Date:   Wed Mar 9 12:52:53 2016 +0100

	    drm/i915: Fix races on fbdev

Committed without R-b by Daniel, so maybe he has a comment on it too.

Regards, Joonas

Quoted 2 lines

-Chris

Signature · 7 lines
Joonas Lahtinen
Open Source Technology Center
Intel Corporation
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#16 lore raw

Re: [PATCH v4 1/2] drm/i915: Protect fbdev across slow or failed initialisation

In reply to Chris Wilson

Hi Chris,

Quoted 14 lines · Chris Wilson
Quoted 12 lines · Joonas Lahtinen

Review

Quoted 9 lines · Chris Wilson

 void intel_fbdev_restore_mode(struct drm_device *dev) {

drivers/gpu/drm/i915/intel_fbdev.c +1 -4
-	int ret;
-	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct intel_fbdev *ifbdev = dev_priv->fbdev;
-	struct drm_fb_helper *fb_helper;
+	struct intel_fbdev *ifbdev = intel_fbdev_get_if_active(dev);

(quoted diff) +0 -1
-	async_synchronize_full();

What's with the async_synchronize_full() begin removed completely?

Quoted 6 lines

Because it's not just wrong, but completely broken imo.

During suspend, we want to freeze the async task not flush. Then here and during resume we skip the restoration of the unregistered fbdev, and then when the task is woken it can complete the registration and present the vanilla ifbdev.

No. The fbdev initialization is guaranteed to have finished before suspend. So "we want to freeze the async task" is wrong thinking. There is no async task to freeze.

Please read the commit message of a7442b93cf32 ("drm/i915: Fix races on fbdev"):

"a device is never suspended until its ->probe callback (and all asynchronous tasks it scheduled) have finished. See dpm_prepare(), which calls wait_for_device_probe(), which calls async_synchronize_full()."

Best regards,

Lukas

Quoted 10 lines

During hibernation, we really just want to cancel the task and start from scratch on resume.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#20 lore raw
In reply to Tomi Sarvela

Hi Tomi,

Quoted 7 lines · Tomi Sarvela

The problem with the results in your link is that there is no HSW, ILK, IVB or SNB results. This might give the impression that everything is well.

Most damning is lack of HSW-gt2 and SNB-dellxps: those machines hang on to APC, and have run quite stably for every Patchwork run. The case isn't strong enough yet that series should fail if either of those won't run, but it might be so in future.

So my patch seeking to fix the hangs has passed Romania CI with "success":
https://patchwork.freedesktop.org/series/5125/

However I don't see HSW-gt2 and SNB-dellxps in their hardware lineup. And I would still like to know what the actual cause of the hangs is since they do not occur on my IVB laptop.

If you get the chance maybe you can repeat the test and include a "dump_stack();" at the beginning of intel_fbdev_output_poll_changed() and intel_fbdev_restore_mode(). This should show in the logs which of the two functions is called during suspend/resume and from where it's called. My guess is that this particular hardware causes a hotplug signal to be generated upon waking up. If the hangs do not stop with this patch, booting with "no_console_suspend" should at least show what's going on.

Thank you & best regards,

Lukas
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#21 lore raw

Hi Lukas,

Ran this patch through the farm, and it seems that this patch might've helped HSW, maybe even BSW.

ILK, IVB and SNB are still hanging hard to the same igt-test, kms_pipe_crc_basic@suspend-read-crc-pipe-?

I'll try to make a kernel with the changes you proposed, but as I'm not familiar with the driver innards, it might take a while.

Tomi

Quoted 28 lines · On Thursday 31 March 2016 22:35:17 Lukas Wunner wrote:

Hi Tomi,

Quoted 8 lines · Tomi Sarvela

The problem with the results in your link is that there is no HSW, ILK, IVB
or SNB results. This might give the impression that everything is well.

Most damning is lack of HSW-gt2 and SNB-dellxps: those machines hang on to APC, and have run quite stably for every Patchwork run. The case isn't strong enough yet that series should fail if either of those won't run, but it might be so in future.

So my patch seeking to fix the hangs has passed Romania CI with "success":
https://patchwork.freedesktop.org/series/5125/

However I don't see HSW-gt2 and SNB-dellxps in their hardware lineup. And I would still like to know what the actual cause of the hangs is since they do not occur on my IVB laptop.

If you get the chance maybe you can repeat the test and include a "dump_stack();" at the beginning of intel_fbdev_output_poll_changed() and intel_fbdev_restore_mode(). This should show in the logs which of the two functions is called during suspend/resume and from where it's called. My guess is that this particular hardware causes a hotplug signal to be generated upon waking up. If the hangs do not stop with this patch, booting with "no_console_suspend" should at least show what's going on.

Quoted 3 lines

Thank you & best regards,

Lukas

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx