Skip to contentSkip to conversation

simplefb: fix unmapping fb during destruction

PATCH In review David Herrmann 3 participants 7 messages Oct 2, 2013 – Oct 30, 2013 Framebuffer layer

lore
Share & tools
lore
#0 lore raw

PATCH 1/2 simplefb: fix unmapping fb during destruction

Unfortunately, fbdev does not create its own "struct device" for framebuffers. Instead, it attaches to the device of the parent layer. This has the side-effect that devm_* managed resources are not cleaned up on framebuffer-destruction but rather during destruction of the parent-device. In case of fbdev this might be too late, though. remove_conflicting_framebuffer() may remove fbdev devices but keep the parent device as it is.

Therefore, we now use plain ioremap() and unmap the framebuffer in the fb_destroy() callback. Note that we must not free the device here as this might race with the parent-device removal. Instead, we rely on unregister_framebuffer() as barrier and we're safe.

---
Hi

I know that simplefb was supposed to stay "as simple as possible" but I really think this series is the last set of fixes I have. Unfortunately framebuffer DRM handover is mandatory so we cannot ignore it in simplefb.

Both patches are not critical at all and are targeted at 3.13-rc1.

Thanks
David

 drivers/video/simplefb.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)
drivers/video/simplefb.c +10 -2
index 8d78106..74b016c 100644
@@ -66,8 +66,15 @@ static int simplefb_setcolreg(u_int regno, u_int red, u_int green, u_int blue,
 	return 0;
 }
 
+static void simplefb_destroy(struct fb_info *info)
+{
+	if (info->screen_base)
+		iounmap(info->screen_base);
+}
+
 static struct fb_ops simplefb_ops = {
 	.owner		= THIS_MODULE,
+	.fb_destroy	= simplefb_destroy,
 	.fb_setcolreg	= simplefb_setcolreg,
 	.fb_fillrect	= cfb_fillrect,
 	.fb_copyarea	= cfb_copyarea,
@@ -212,8 +219,8 @@ static int simplefb_probe(struct platform_device *pdev)
 
 	info->fbops = &simplefb_ops;
 	info->flags = FBINFO_DEFAULT | FBINFO_MISC_FIRMWARE;
-	info->screen_base = devm_ioremap(&pdev->dev, info->fix.smem_start,
-					 info->fix.smem_len);
+	info->screen_base = ioremap(info->fix.smem_start,
+				    info->fix.smem_len);
 	if (!info->screen_base) {
 		framebuffer_release(info);
 		return -ENODEV;
@@ -223,6 +230,7 @@ static int simplefb_probe(struct platform_device *pdev)
 	ret = register_framebuffer(info);
 	if (ret < 0) {
 		dev_err(&pdev->dev, "Unable to register simplefb: %d\n", ret);
+		iounmap(info->screen_base);
 		framebuffer_release(info);
 		return ret;
 	}
Signature · 1 line
1.8.4
#1 lore raw

PATCH 2/2 simplefb: use write-combined remapping

Framebuffers shouldn't be cached and it is usually very uncommon to read them. Therefore, use ioremap_wc() to get significant speed improvements on systems which provide it. On all other systems it's aliased to ioremap_nocache() which is also fine.

---
 drivers/video/simplefb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/video/simplefb.c +2 -2
index 74b016c..64db54a 100644
@@ -219,8 +219,8 @@ static int simplefb_probe(struct platform_device *pdev)
 
 	info->fbops = &simplefb_ops;
 	info->flags = FBINFO_DEFAULT | FBINFO_MISC_FIRMWARE;
-	info->screen_base = ioremap(info->fix.smem_start,
-				    info->fix.smem_len);
+	info->screen_base = ioremap_wc(info->fix.smem_start,
+				       info->fix.smem_len);
 	if (!info->screen_base) {
 		framebuffer_release(info);
 		return -ENODEV;
Signature · 1 line
1.8.4
#2 lore raw

Re: [PATCH 2/2] simplefb: use write-combined remapping

Hi Tomi

Ping?

Thanks
David

Quoted 32 lines · David Herrmann

Framebuffers shouldn't be cached and it is usually very uncommon to read them. Therefore, use ioremap_wc() to get significant speed improvements on systems which provide it. On all other systems it's aliased to ioremap_nocache() which is also fine.

---
 drivers/video/simplefb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/video/simplefb.c +2 -3
index 74b016c..64db54a 100644
@@ -219,8 +219,8 @@ static int simplefb_probe(struct platform_device *pdev)
 
        info->fbops = &simplefb_ops;
        info->flags = FBINFO_DEFAULT | FBINFO_MISC_FIRMWARE;
-       info->screen_base = ioremap(info->fix.smem_start,
-                                   info->fix.smem_len);
+       info->screen_base = ioremap_wc(info->fix.smem_start,
+                                      info->fix.smem_len);
        if (!info->screen_base) {
                framebuffer_release(info);
                return -ENODEV;
--

1.8.4

#3 lore raw

Re: [PATCH 2/2] simplefb: use write-combined remapping

Quoted 3 lines · David Herrmann

Hi Tomi

Ping?

Thanks, queued this and the 1/2 patch for 3.13.

Tomi

Quoted 37 lines

Thanks
David

Quoted 32 lines · David Herrmann

Framebuffers shouldn't be cached and it is usually very uncommon to read them. Therefore, use ioremap_wc() to get significant speed improvements on systems which provide it. On all other systems it's aliased to ioremap_nocache() which is also fine.

---
 drivers/video/simplefb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/video/simplefb.c +2 -3
index 74b016c..64db54a 100644
@@ -219,8 +219,8 @@ static int simplefb_probe(struct platform_device *pdev)
 
        info->fbops = &simplefb_ops;
        info->flags = FBINFO_DEFAULT | FBINFO_MISC_FIRMWARE;
-       info->screen_base = ioremap(info->fix.smem_start,
-                                   info->fix.smem_len);
+       info->screen_base = ioremap_wc(info->fix.smem_start,
+                                      info->fix.smem_len);
        if (!info->screen_base) {
                framebuffer_release(info);
                return -ENODEV;
--

1.8.4

Acked-by on v1 1/2
#4 lore raw
In reply to David Herrmann
Quoted 8 lines · David Herrmann

Unfortunately, fbdev does not create its own "struct device" for framebuffers. Instead, it attaches to the device of the parent layer. This has the side-effect that devm_* managed resources are not cleaned up on framebuffer-destruction but rather during destruction of the parent-device. In case of fbdev this might be too late, though. remove_conflicting_framebuffer() may remove fbdev devices but keep the parent device as it is.

Quoted 4 lines

Therefore, we now use plain ioremap() and unmap the framebuffer in the fb_destroy() callback. Note that we must not free the device here as this might race with the parent-device removal. Instead, we rely on unregister_framebuffer() as barrier and we're safe.

So, once the .fb_destroy callback has been executed, there's no other callback to resurrect it? The framebuffer itself is still registered until the device's remove, yet after .fb_destroy, the memory is unmapped, which would be dangerous if the FB can be re-started.

If that's not an issue, this patch seems fine, so

Quoted 3 lines

I know that simplefb was supposed to stay "as simple as possible" but I really think this series is the last set of fixes I have. Unfortunately framebuffer DRM handover is mandatory so we cannot ignore it in simplefb.

I don't think this patch adds any significant complexity, so I'm not worried at least.

#5 lore raw

Hi

Quoted 14 lines · Stephen Warren
Quoted 8 lines · David Herrmann

Unfortunately, fbdev does not create its own "struct device" for framebuffers. Instead, it attaches to the device of the parent layer. This has the side-effect that devm_* managed resources are not cleaned up on framebuffer-destruction but rather during destruction of the parent-device. In case of fbdev this might be too late, though. remove_conflicting_framebuffer() may remove fbdev devices but keep the parent device as it is.

Quoted 4 lines

Therefore, we now use plain ioremap() and unmap the framebuffer in the fb_destroy() callback. Note that we must not free the device here as this might race with the parent-device removal. Instead, we rely on unregister_framebuffer() as barrier and we're safe.

Quoted 4 lines

So, once the .fb_destroy callback has been executed, there's no other callback to resurrect it? The framebuffer itself is still registered until the device's remove, yet after .fb_destroy, the memory is unmapped, which would be dangerous if the FB can be re-started.

fbdev lifetime tracking is weird.. ->fb_destroy() is called by unregister_framebuffer() _after_ it got unlinked. So no, the framebuffer is gone and cannot be resurrected. However, the unregistered/dead fbdev object itself stays around until you call framebuffer_release(). We cannot call it from fb_destroy(), though as the platform_data in your platform device still points to the fbdev device. Therefore we keep it and wait for the platform-driver to be removed which then again calls unregister_framebuffer() (which will immediately return as the fbdev device is not registered) and then you can finally call framebuffer_release().

Note that even though there's fb_get_info() and fb_put_info(), both are not exported and never used. So there is *no* fbdev ref-counting (which would be horribly broken anyway) and the driver is the sole owner of the fbdev object.

Quoted 2 lines

If that's not an issue, this patch seems fine, so

Thanks!

Quoted 6 lines
Quoted 3 lines

I know that simplefb was supposed to stay "as simple as possible" but I really think this series is the last set of fixes I have. Unfortunately framebuffer DRM handover is mandatory so we cannot ignore it in simplefb.

I don't think this patch adds any significant complexity, so I'm not worried at least.

Good to hear!

Thanks
David

#6 lore raw

Hi Tomi

Could we get this in -next before the merge-window starts? Stephen already ack'ed it.

Thanks
David

Quoted 54 lines · David Herrmann

Hi

Quoted 14 lines · Stephen Warren
Quoted 8 lines · David Herrmann

Unfortunately, fbdev does not create its own "struct device" for framebuffers. Instead, it attaches to the device of the parent layer. This has the side-effect that devm_* managed resources are not cleaned up on framebuffer-destruction but rather during destruction of the parent-device. In case of fbdev this might be too late, though. remove_conflicting_framebuffer() may remove fbdev devices but keep the parent device as it is.

Quoted 4 lines

Therefore, we now use plain ioremap() and unmap the framebuffer in the fb_destroy() callback. Note that we must not free the device here as this might race with the parent-device removal. Instead, we rely on unregister_framebuffer() as barrier and we're safe.

Quoted 4 lines

So, once the .fb_destroy callback has been executed, there's no other callback to resurrect it? The framebuffer itself is still registered until the device's remove, yet after .fb_destroy, the memory is unmapped, which would be dangerous if the FB can be re-started.

fbdev lifetime tracking is weird.. ->fb_destroy() is called by unregister_framebuffer() _after_ it got unlinked. So no, the framebuffer is gone and cannot be resurrected. However, the unregistered/dead fbdev object itself stays around until you call framebuffer_release(). We cannot call it from fb_destroy(), though as the platform_data in your platform device still points to the fbdev device. Therefore we keep it and wait for the platform-driver to be removed which then again calls unregister_framebuffer() (which will immediately return as the fbdev device is not registered) and then you can finally call framebuffer_release().

Note that even though there's fb_get_info() and fb_put_info(), both are not exported and never used. So there is *no* fbdev ref-counting (which would be horribly broken anyway) and the driver is the sole owner of the fbdev object.

Quoted 2 lines

If that's not an issue, this patch seems fine, so

Thanks!

Quoted 6 lines
Quoted 3 lines

I know that simplefb was supposed to stay "as simple as possible" but I really think this series is the last set of fixes I have. Unfortunately framebuffer DRM handover is mandatory so we cannot ignore it in simplefb.

I don't think this patch adds any significant complexity, so I'm not worried at least.

Good to hear!

Thanks
David