Skip to contentSkip to conversation

drm/vmwgfx: Stop requesting the pci regions

PATCH In review Zack Rusin 4 participants 18 messages Jan 17, 2022 – Jan 21, 2022

Subsystems (8)
lore
Share & tools
lore

Zack Rusin

author
#0 lore raw

PATCH 1/1 drm/vmwgfx: Stop requesting the pci regions

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions are held by the platform. In that case remove_conflicting*_framebuffers only removes the simplefb but not the regions held by sysfb.

Like the other drm drivers we need to stop requesting all the pci regions to let the driver load with platform code enabled.
This allows vmwgfx to load correctly on systems with sysfb_simple enabled.

Fixes 523375c943e5 ("drm/vmwgfx: Port vmwgfx to arm64")
---
 drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 4 ----
 1 file changed, 4 deletions(-)
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +0 -4
index fe36efdb7ff5..27feb19f3324 100644
@@ -724,10 +724,6 @@ static int vmw_setup_pci_resources(struct vmw_private *dev,
 
 	pci_set_master(pdev);
 
-	ret = pci_request_regions(pdev, "vmwgfx probe");
-	if (ret)
-		return ret;
-
 	dev->pci_id = pci_id;
 	if (pci_id == VMWGFX_PCI_ID_SVGA3) {
 		rmmio_start = pci_resource_start(pdev, 0);
Signature · 1 line
2.32.0
maintainer Reviewed-by on v1 1/1
#1 lore raw

Hello Zack,

Quoted 6 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions are held by the platform. In that case remove_conflicting*_framebuffers only removes the simplefb but not the regions held by sysfb.

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY flag from the memory resource added to the "simple-framebuffer" device ?

In fact, maybe in sysfb_create_simplefb() shouldn't even attempt to claim a memory resource and just register the platform device with no resources ?

Quoted 4 lines

Like the other drm drivers we need to stop requesting all the pci regions to let the driver load with platform code enabled.
This allows vmwgfx to load correctly on systems with sysfb_simple enabled.

I read this very interesting thread from two years ago:

https://lkml.org/lkml/2020/11/5/248

Maybe is worth mentioning in the commit message what Daniel said there, that is that only a few DRM drivers request explicitly the PCI regions and the only reliable approach is for bus drivers to claim these.

In other words, removing the pci_request_regions() seems to have merit on its own.

Quoted 6 lines
Fixes 523375c943e5 ("drm/vmwgfx: Port vmwgfx to arm64")

---

The patch looks good to me, thanks a lot for fixing this:

Best regards,

Signature · 3 lines
Javier Martinez Canillas
Linux Engineering
Red Hat
#2 lore raw
Quoted 10 lines · Javier Martinez Canillas

Hello Zack,

Quoted 6 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions are held by the platform. In that case remove_conflicting*_framebuffers only removes the simplefb but not the regions held by sysfb.

Quoted 6 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY flag from the memory resource added to the "simple-framebuffer" device ?

In fact, maybe in sysfb_create_simplefb() shouldn't even attempt to claim a memory resource and just register the platform device with no resources ?

Answering my own question, this would mean adding that platform specific logic to the drivers matching the "simple-framebuffer" device so it should remain in sysfb_create_simplefb().

But I still wonder if dropping the IORESOURCE_BUSY flag is something that will be reasonable to prevent other drivers having the the issue reported in this patch.

Best regards,

Signature · 3 lines
Javier Martinez Canillas
Linux Engineering
Red Hat
#3 lore raw
In reply to Javier Martinez Canillas
Quoted 11 lines · Javier Martinez Canillas

Hello Zack,

Quoted 7 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 3 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY flag from the memory resource added to the "simple-framebuffer" device ?

I think this is one of those cases where it depends on what we plan to do after that. Sementically it makes sense to have it in there - the framebuffer memory is claimed by the "simple-framebuffer" and it's busy, it's just that it creates issues for drivers after unloading. I think removing it, while making things easier for drivers, would be confusing for people reading the code later, unless there's some kind of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better solution than having the drm drivers not call pci_request_region[s], which apart from hyperv and cirrus (iirc bochs does it for resources other than fb which wouldn't have been claimed by "simple-framebuffer") is already the case.

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove pci_request_region[s] from hyperv and cirrus.

Quoted 7 lines
Quoted 6 lines

Like the other drm drivers we need to stop requesting all the pci regions
to let the driver load with platform code enabled.
This allows vmwgfx to load correctly on systems with sysfb_simple enabled.

Quoted 7 lines

I read this very interesting thread from two years ago:

https://lkml.org/lkml/2020/11/5/248

Maybe is worth mentioning in the commit message what Daniel said there, that is that only a few DRM drivers request explicitly the PCI regions and the only reliable approach is for bus drivers to claim these.

Ah, great point. I'll update the commit log with that.

Quoted 10 lines
Quoted 6 lines
Fixes 523375c943e5 ("drm/vmwgfx: Port vmwgfx to arm64")

---

The patch looks good to me, thanks a lot for fixing this:

Thanks for taking a look at this, I appreciate it!

z

#4 lore raw

Hi

Am 19.01.22 um 03:15 schrieb Zack Rusin:

Quoted 30 lines · Javier Martinez Canillas
Quoted 11 lines

Hello Zack,

Quoted 7 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 3 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY flag from the memory resource added to the "simple-framebuffer" device ?

I think this is one of those cases where it depends on what we plan to do after that. Sementically it makes sense to have it in there - the framebuffer memory is claimed by the "simple-framebuffer" and it's busy, it's just that it creates issues for drivers after unloading. I think removing it, while making things easier for drivers, would be confusing for people reading the code later, unless there's some kind of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better solution than having the drm drivers not call pci_request_region[s], which apart from hyperv and cirrus (iirc bochs does it for resources other than fb which wouldn't have been claimed by "simple-framebuffer") is already the case.

Quoted 3 lines

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove pci_request_region[s] from hyperv and cirrus.

I just discussed this a bit with Javier. It's a problem with the simple-framebuffer code, rather then vmwgfx.

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have drivers register/release the range with _BUSY. That would signal the memory belongs to the sysfb device but is not busy unless a driver has been bound. After simplefb released the range, it should be 'non-busy' again and available for vmwgfx. Simpledrm does a hot-unplug of the sysfb device, so the memory range gets released entirely. If you want, I'll prepare some patches for this scenario.

If this doesn't work, pushing all request/release pairs into drivers would be my next option.

If none of this is feasible, we can still remove pci_request_region() from vmwgfx.

Best regards
Thomas

Quoted 34 lines
Quoted 10 lines
Quoted 9 lines

Like the other drm drivers we need to stop requesting all the pci regions
to let the driver load with platform code enabled.
This allows vmwgfx to load correctly on systems with sysfb_simple enabled.

Quoted 7 lines

I read this very interesting thread from two years ago:

https://lkml.org/lkml/2020/11/5/248

Maybe is worth mentioning in the commit message what Daniel said there, that is that only a few DRM drivers request explicitly the PCI regions and the only reliable approach is for bus drivers to claim these.

Ah, great point. I'll update the commit log with that.

Quoted 10 lines
Quoted 6 lines
Fixes 523375c943e5 ("drm/vmwgfx: Port vmwgfx to arm64")

---

The patch looks good to me, thanks a lot for fixing this:

Thanks for taking a look at this, I appreciate it!

z

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#5 lore raw

Hi Zack

Am 19.01.22 um 10:13 schrieb Thomas Zimmermann:

Quoted 33 lines

Hi

Am 19.01.22 um 03:15 schrieb Zack Rusin:

Quoted 30 lines · Javier Martinez Canillas
Quoted 11 lines

Hello Zack,

Quoted 7 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 3 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY flag from the memory resource added to the "simple-framebuffer" device ?

I think this is one of those cases where it depends on what we plan to do after that. Sementically it makes sense to have it in there - the framebuffer memory is claimed by the "simple-framebuffer" and it's busy, it's just that it creates issues for drivers after unloading. I think removing it, while making things easier for drivers, would be confusing for people reading the code later, unless there's some kind of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better solution than having the drm drivers not call pci_request_region[s], which apart from hyperv and cirrus (iirc bochs does it for resources other than fb which wouldn't have been claimed by "simple-framebuffer") is already the case.

Quoted 7 lines
Quoted 3 lines

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove pci_request_region[s] from hyperv and cirrus.

I just discussed this a bit with Javier. It's a problem with the simple-framebuffer code, rather then vmwgfx.

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have drivers register/release the range with _BUSY. That would signal the memory belongs to the sysfb device but is not busy unless a driver has been bound. After simplefb released the range, it should be 'non-busy' again and available for vmwgfx. Simpledrm does a hot-unplug of the sysfb device, so the memory range gets released entirely. If you want, I'll prepare some patches for this scenario.

Attached is a patch that implements this. Doing

cat /proc/iomem
...
e0000000-efffffff : 0000:00:02.0

    e0000000-e07e8fff : BOOTFB

      e0000000-e07e8fff : simplefb

...

shows the memory. 'BOOTFB' is the simple-framebuffer device and 'simplefb' is the driver. Only the latter uses _BUSY. Same for simpledrm. Once the drivers have been unloaded, the BUSY flag is gone and the memory canbe acquired by vmwgfx.

Zack, please test this patch. If it works, I'll send out the real patchset.

Best regards
Thomas

Quoted 43 lines

If this doesn't work, pushing all request/release pairs into drivers would be my next option.

If none of this is feasible, we can still remove pci_request_region() from vmwgfx.

Best regards
Thomas

Quoted 34 lines
Quoted 10 lines
Quoted 9 lines

Like the other drm drivers we need to stop requesting all the pci regions
to let the driver load with platform code enabled.
This allows vmwgfx to load correctly on systems with sysfb_simple enabled.

Quoted 7 lines

I read this very interesting thread from two years ago:

https://lkml.org/lkml/2020/11/5/248

Maybe is worth mentioning in the commit message what Daniel said there, that is that only a few DRM drivers request explicitly the PCI regions and the only reliable approach is for bus drivers to claim these.

Ah, great point. I'll update the commit log with that.

Quoted 10 lines
Quoted 6 lines
Fixes 523375c943e5 ("drm/vmwgfx: Port vmwgfx to arm64")

---

The patch looks good to me, thanks a lot for fixing this:

Thanks for taking a look at this, I appreciate it!

z

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

From f5da496296f7dcaf8754e1b0446660b100539a73 Mon Sep 17 00:00:00 2001
From: Thomas Zimmermann <[email protected]>
Date: Wed, 19 Jan 2022 10:23:07 +0100
Subject: [PATCH] [RFC] drop IORESOURCE_BUSY from sysfb code and request memory
in drivers

Drop the IORESOURCE_BUSY flag when creating a simple-framebuffer device. Instead acquire the memory in drivers.

---
 drivers/firmware/sysfb_simplefb.c |  2 +-
 drivers/gpu/drm/tiny/simpledrm.c  | 17 ++++++---
 drivers/video/fbdev/simplefb.c    | 57 ++++++++++++++++++++++---------
 3 files changed, 53 insertions(+), 23 deletions(-)
drivers/firmware/sysfb_simplefb.c +1 -1
index b86761904949..179e9d0ef3e9 100644
@@ -99,7 +99,7 @@ __init int sysfb_create_simplefb(const struct screen_info *si,
 
 	/* setup IORESOURCE_MEM as framebuffer memory */
 	memset(&res, 0, sizeof(res));
-	res.flags = IORESOURCE_MEM | IORESOURCE_BUSY;
+	res.flags = IORESOURCE_MEM;
 	res.name = simplefb_resname;
 	res.start = base;
 	res.end = res.start + length - 1;
drivers/gpu/drm/tiny/simpledrm.c +12 -5
index 04146da2d1d8..32550697b1a7 100644
@@ -526,21 +526,28 @@ static int simpledrm_device_init_mm(struct simpledrm_device *sdev)
 {
 	struct drm_device *dev = &sdev->dev;
 	struct platform_device *pdev = sdev->pdev;
-	struct resource *mem;
+	struct resource *res, *mem;
 	void __iomem *screen_base;
 	int ret;
 
-	mem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
-	if (!mem)
+	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	if (!res)
 		return -EINVAL;
 
-	ret = devm_aperture_acquire_from_firmware(dev, mem->start, resource_size(mem));
+	ret = devm_aperture_acquire_from_firmware(dev, res->start, resource_size(res));
 	if (ret) {
 		drm_err(dev, "could not acquire memory range %pr: error %d\n",
-			mem, ret);
+			res, ret);
 		return ret;
 	}
 
+	mem = devm_request_mem_region(&pdev->dev, res->start, resource_size(res),
+				      sdev->dev.driver->name);
+	if (!mem) {
+		drm_err(dev, "could not acquire memory region %pr\n", res);
+		return -EBUSY;
+	}
+
 	screen_base = devm_ioremap_wc(&pdev->dev, mem->start,
 				      resource_size(mem));
 	if (!screen_base)
drivers/video/fbdev/simplefb.c +40 -17

103 lines not on this page · raw message

Signature · 1 line
2.34.1
#6 lore raw
Quoted 74 lines · Thomas Zimmermann

Hi Zack

Am 19.01.22 um 10:13 schrieb Thomas Zimmermann:

Quoted 50 lines

Hi

Am 19.01.22 um 03:15 schrieb Zack Rusin:

Quoted 40 lines · Javier Martinez Canillas
Quoted 12 lines

Hello Zack,

Quoted 8 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions
are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 5 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY
flag from the memory resource added to the "simple-framebuffer" device
?

I think this is one of those cases where it depends on what we plan to
do after that. Sementically it makes sense to have it in there - the
framebuffer memory is claimed by the "simple-framebuffer" and it's busy, it's just that it creates issues for drivers after unloading. I
think removing it, while making things easier for drivers, would be confusing for people reading the code later, unless there's some kind
of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better solution than having the drm drivers not call pci_request_region[s],
which apart from hyperv and cirrus (iirc bochs does it for resources
other than fb which wouldn't have been claimed by "simple- framebuffer")
is already the case.

Quoted 3 lines

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove pci_request_region[s] from hyperv and cirrus.

I just discussed this a bit with Javier. It's a problem with the simple-framebuffer code, rather then vmwgfx.

Quoted 10 lines

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have drivers register/release the range with _BUSY. That would signal the memory belongs to the sysfb device but is not busy unless a driver has
been bound. After simplefb released the range, it should be 'non- busy'
again and available for vmwgfx. Simpledrm does a hot-unplug of the sysfb
device, so the memory range gets released entirely. If you want, I'll prepare some patches for this scenario.

Attached is a patch that implements this. Doing

  cat /proc/iomem
   ...
   e0000000-efffffff : 0000:00:02.0

     e0000000-e07e8fff : BOOTFB

       e0000000-e07e8fff : simplefb

Quoted 8 lines

   ...

shows the memory. 'BOOTFB' is the simple-framebuffer device and 'simplefb' is the driver. Only the latter uses _BUSY. Same for and the memory canbe acquired by vmwgfx.

Zack, please test this patch. If it works, I'll send out the real patchset.

Hmm, the patch looks good but it doesn't work. After boot: /proc/iomem 50000000-7fffffff : pcie@0x40000000
78000000-7fffffff : 0000:00:0f.0

    78000000-782fffff : BOOTFB

and vmwgfx fails on pci_request_regions:

kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

leaving the system without a fb driver.

z

#7 lore raw

Hi

Am 19.01.22 um 16:12 schrieb Zack Rusin:

Quoted 91 lines · Thomas Zimmermann
Quoted 74 lines

Hi Zack

Am 19.01.22 um 10:13 schrieb Thomas Zimmermann:

Quoted 50 lines

Hi

Am 19.01.22 um 03:15 schrieb Zack Rusin:

Quoted 40 lines · Javier Martinez Canillas
Quoted 12 lines

Hello Zack,

Quoted 8 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions
are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 5 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY
flag from the memory resource added to the "simple-framebuffer" device
?

I think this is one of those cases where it depends on what we plan to
do after that. Sementically it makes sense to have it in there - the
framebuffer memory is claimed by the "simple-framebuffer" and it's busy, it's just that it creates issues for drivers after unloading. I
think removing it, while making things easier for drivers, would be confusing for people reading the code later, unless there's some kind
of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better solution than having the drm drivers not call pci_request_region[s],
which apart from hyperv and cirrus (iirc bochs does it for resources
other than fb which wouldn't have been claimed by "simple- framebuffer")
is already the case.

Quoted 3 lines

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove pci_request_region[s] from hyperv and cirrus.

I just discussed this a bit with Javier. It's a problem with the simple-framebuffer code, rather then vmwgfx.

Quoted 10 lines

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have drivers register/release the range with _BUSY. That would signal the memory belongs to the sysfb device but is not busy unless a driver has
been bound. After simplefb released the range, it should be 'non- busy'
again and available for vmwgfx. Simpledrm does a hot-unplug of the sysfb
device, so the memory range gets released entirely. If you want, I'll prepare some patches for this scenario.

Attached is a patch that implements this. Doing

  cat /proc/iomem
   ...
   e0000000-efffffff : 0000:00:02.0

     e0000000-e07e8fff : BOOTFB

       e0000000-e07e8fff : simplefb

Quoted 8 lines

   ...

shows the memory. 'BOOTFB' is the simple-framebuffer device and 'simplefb' is the driver. Only the latter uses _BUSY. Same for and the memory canbe acquired by vmwgfx.

Zack, please test this patch. If it works, I'll send out the real patchset.

Hmm, the patch looks good but it doesn't work. After boot: /proc/iomem 50000000-7fffffff : pcie@0x40000000
78000000-7fffffff : 0000:00:0f.0

     78000000-782fffff : BOOTFB

and vmwgfx fails on pci_request_regions:

Quoted 7 lines

kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

leaving the system without a fb driver.

OK, I suspect that it would work if you use simpledrm instead of simplefb. Could you try please? You'd have to build DRM and simpledrm into the kernel binary.

If that works, would you consider protecting pci_request_region() with
#if not defined(CONFIG_FB_SIMPLE)
#endif

with a FIXME comment?

simpledrm hot-unplugs the sysfb device entirely. Maybe simplefb can do the same. I'm not sure, but possibly.

Best regards
Thomas

Quoted 2 lines

z

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#8 lore raw

On 1/19/22 16:50, Thomas Zimmermann wrote:

[snip]

Quoted 43 lines
Quoted 43 lines
Quoted 21 lines
Quoted 10 lines

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have drivers register/release the range with _BUSY. That would signal the memory belongs to the sysfb device but is not busy unless a driver has
been bound. After simplefb released the range, it should be 'non- busy'
again and available for vmwgfx. Simpledrm does a hot-unplug of the sysfb
device, so the memory range gets released entirely. If you want, I'll prepare some patches for this scenario.

Attached is a patch that implements this. Doing

  cat /proc/iomem
   ...
   e0000000-efffffff : 0000:00:02.0

     e0000000-e07e8fff : BOOTFB

       e0000000-e07e8fff : simplefb

Quoted 8 lines

   ...

shows the memory. 'BOOTFB' is the simple-framebuffer device and 'simplefb' is the driver. Only the latter uses _BUSY. Same for and the memory canbe acquired by vmwgfx.

Zack, please test this patch. If it works, I'll send out the real patchset.

Hmm, the patch looks good but it doesn't work. After boot: /proc/iomem 50000000-7fffffff : pcie@0x40000000
78000000-7fffffff : 0000:00:0f.0

     78000000-782fffff : BOOTFB

and vmwgfx fails on pci_request_regions:

kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

Quoted 6 lines
Quoted 1 line

leaving the system without a fb driver.

OK, I suspect that it would work if you use simpledrm instead of simplefb. Could you try please? You'd have to build DRM and simpledrm into the kernel binary.

Yes, I believe that should work.
Zack, could you please try if just the following [0] make it works ?

That is, dropping the IORESOURCE_BUSY but not doing the memory region request / release in simplefb and keeping it in the vmwgfx driver.

[0]:
From da6de1430b9dc252eccf2d6fee0446d33375fa6d Mon Sep 17 00:00:00 2001
From: Javier Martinez Canillas <[email protected]>
Date: Wed, 19 Jan 2022 14:41:25 +0100
Subject: [PATCH] drivers/firmware: Don't mark as busy the simple-framebuffer
IO resource

The sysfb_create_simplefb() function requests a IO memory resource for the simple-framebuffer platform device, but it also marks it as busy which led to drivers requesting the same memory resource to fail.

Let's drop the IORESOURCE_BUSY flag and let drivers to request it as busy instead.

---
 drivers/firmware/sysfb_simplefb.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/firmware/sysfb_simplefb.c +1 -1
index 303a491e520d..76c4abc42a30 100644
@@ -99,7 +99,7 @@ __init int sysfb_create_simplefb(const struct screen_info *si,
 
 	/* setup IORESOURCE_MEM as framebuffer memory */
 	memset(&res, 0, sizeof(res));
-	res.flags = IORESOURCE_MEM | IORESOURCE_BUSY;
+	res.flags = IORESOURCE_MEM;
 	res.name = simplefb_resname;
 	res.start = base;
 	res.end = res.start + length - 1;
Signature · 1 line
2.33.1
#9 lore raw
Quoted 56 lines · Javier Martinez Canillas

On 1/19/22 16:50, Thomas Zimmermann wrote:

[snip]

Quoted 52 lines
Quoted 51 lines
Quoted 27 lines
Quoted 8 lines

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have
drivers register/release the range with _BUSY. That would signal the
memory belongs to the sysfb device but is not busy unless a driver
has
been bound. After simplefb released the range, it should be

Quoted 8 lines

'non-
busy'
again and available for vmwgfx. Simpledrm does a hot-unplug of the
sysfb
device, so the memory range gets released entirely. If you want, I'll
prepare some patches for this scenario.

Attached is a patch that implements this. Doing

   cat /proc/iomem
    ...
    e0000000-efffffff : 0000:00:02.0

      e0000000-e07e8fff : BOOTFB

        e0000000-e07e8fff : simplefb

Quoted 8 lines

    ...

shows the memory. 'BOOTFB' is the simple-framebuffer device and 'simplefb' is the driver. Only the latter uses _BUSY. Same for and the memory canbe acquired by vmwgfx.

Zack, please test this patch. If it works, I'll send out the real patchset.

Hmm, the patch looks good but it doesn't work. After boot:
/proc/iomem
50000000-7fffffff : pcie@0x40000000 78000000-7fffffff : 0000:00:0f.0 78000000-782fffff : BOOTFB

and vmwgfx fails on pci_request_regions:

kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

leaving the system without a fb driver.

Quoted 7 lines
Quoted 4 lines

OK, I suspect that it would work if you use simpledrm instead of simplefb. Could you try please? You'd have to build DRM and simpledrm into the kernel binary.

Yes, I believe that should work.
 Zack, could you please try if just the following [0] make it works ?

Unfortunately that still fails with the same:
kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

z

#10 lore raw
In reply to Thomas Zimmermann
Quoted 122 lines · Thomas Zimmermann

Hi

Am 19.01.22 um 16:12 schrieb Zack Rusin:

Quoted 119 lines · Thomas Zimmermann
Quoted 95 lines

Hi Zack

Am 19.01.22 um 10:13 schrieb Thomas Zimmermann:

Quoted 65 lines

Hi

Am 19.01.22 um 03:15 schrieb Zack Rusin:

Quoted 52 lines
Quoted 14 lines · Javier Martinez Canillas

Hello Zack,

Quoted 10 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the
regions
are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 6 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY
flag from the memory resource added to the "simple- framebuffer"
device
?

I think this is one of those cases where it depends on what we plan
to
do after that. Sementically it makes sense to have it in there -
the
framebuffer memory is claimed by the "simple-framebuffer" and it's
busy, it's just that it creates issues for drivers after unloading.
I
think removing it, while making things easier for drivers, would be
confusing for people reading the code later, unless there's some
kind
of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY
altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better
solution than having the drm drivers not call pci_request_region[s],
which apart from hyperv and cirrus (iirc bochs does it for resources
other than fb which wouldn't have been claimed by "simple- framebuffer")
is already the case.

Quoted 5 lines

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove
pci_request_region[s] from hyperv and cirrus.

I just discussed this a bit with Javier. It's a problem with the
simple-framebuffer code, rather then vmwgfx.

Quoted 16 lines

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have
drivers register/release the range with _BUSY. That would signal the
memory belongs to the sysfb device but is not busy unless a driver
has
been bound. After simplefb released the range, it should be 'non-
busy'
again and available for vmwgfx. Simpledrm does a hot-unplug of the
sysfb
device, so the memory range gets released entirely. If you want, I'll
prepare some patches for this scenario.

Attached is a patch that implements this. Doing

   cat /proc/iomem
    ...
    e0000000-efffffff : 0000:00:02.0

      e0000000-e07e8fff : BOOTFB

        e0000000-e07e8fff : simplefb

Quoted 8 lines

    ...

shows the memory. 'BOOTFB' is the simple-framebuffer device and 'simplefb' is the driver. Only the latter uses _BUSY. Same for and the memory canbe acquired by vmwgfx.

Zack, please test this patch. If it works, I'll send out the real patchset.

Hmm, the patch looks good but it doesn't work. After boot:
/proc/iomem
50000000-7fffffff : pcie@0x40000000 78000000-7fffffff : 0000:00:0f.0 78000000-782fffff : BOOTFB

and vmwgfx fails on pci_request_regions:

kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

Quoted 4 lines
Quoted 1 line

leaving the system without a fb driver.

OK, I suspect that it would work if you use simpledrm instead of into the kernel binary.

Yes, simpledrm works fine. BTW, is there any remaining work before distros can enable it by default?

Quoted 7 lines

If that works, would you consider protecting pci_request_region() with
  #if not defined(CONFIG_FB_SIMPLE)
  #endif

with a FIXME comment?

Yes, I think that's a good compromise. I'll respin the patch with that.

z

#11 lore raw

Hello Zack,

On 1/20/22 05:06, Zack Rusin wrote:

[snip]

Quoted 15 lines
Quoted 15 lines
Quoted 15 lines

Hmm, the patch looks good but it doesn't work. After boot:
/proc/iomem
50000000-7fffffff : pcie@0x40000000 78000000-7fffffff : 0000:00:0f.0 78000000-782fffff : BOOTFB

and vmwgfx fails on pci_request_regions:

kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

leaving the system without a fb driver.

OK, I suspect that it would work if you use simpledrm instead of into the kernel binary.

Yes, simpledrm works fine. BTW, is there any remaining work before distros can enable it by default?

Alpine already did AFAIK, OpenSUSE and Fedora are doing it soon

I don't know about the others distros but I guess they will follow.

Quoted 10 lines
Quoted 7 lines

If that works, would you consider protecting pci_request_region() with
  #if not defined(CONFIG_FB_SIMPLE)
  #endif

with a FIXME comment?

Yes, I think that's a good compromise. I'll respin the patch with that.

Agreed. Thanks a lot for testing the other patches anyways.

Best regards,

Signature · 3 lines
Javier Martinez Canillas
Linux Engineering
Red Hat
#12 lore raw
In reply to Zack Rusin

Hi

Am 20.01.22 um 05:06 schrieb Zack Rusin:
[...]

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

kernel: fb0: switching to vmwgfx from simple
kernel: Console: switching to colour dummy device 80x25
kernel: vmwgfx 0000:00:0f.0: BAR 2: can't reserve [mem 0x78000000- 0x7fffffff 64bit pref]
kernel: vmwgfx: probe of 0000:00:0f.0 failed with error -16

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

leaving the system without a fb driver.

OK, I suspect that it would work if you use simpledrm instead of into the kernel binary.

Yes, simpledrm works fine. BTW, is there any remaining work before distros can enable it by default?

It's here and ready for use.

Simpledrm requires DRM to be linked in the kernel. We're modularizing DRM helpers to reduce the binary size.

On the distro side, Javier and me are sorting out issue. But it's fairly quite and not many problems show up.

Quoted 10 lines
Quoted 8 lines

If that works, would you consider protecting pci_request_region() with
  #if not defined(CONFIG_FB_SIMPLE)
  #endif

with a FIXME comment?

Yes, I think that's a good compromise. I'll respin the patch with that.

Before you do that, I have one more patch for you to try. It's all the changes as before, but now fbdev hot-unplugs the underlying platform device from the device hierarchy. The BOOTFB will not consume parts of vmwgfx's display memory range any longer. It's now the same behavior as with simpledrm.

This works for me with simplefb and efifb on i915 hardware.

Best regards
Thomas

Quoted 2 lines

z

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

From dfd482bfc143b9ec0d8b103a56f6576acef8a860 Mon Sep 17 00:00:00 2001
From: Thomas Zimmermann <[email protected]>
Date: Wed, 19 Jan 2022 10:23:07 +0100
Subject: [PATCH v2] [RFC] drop IORESOURCE_BUSY from sysfb code and request
memory in drivers

Drop the IORESOURCE_BUSY flag when creating a simple-framebuffer device. Instead acquire the memory in drivers. Also hot-unplug the platform device.

---
 drivers/firmware/sysfb_simplefb.c |  2 +-
 drivers/gpu/drm/tiny/simpledrm.c  | 17 ++++++---
 drivers/video/fbdev/core/fbmem.c  | 29 ++++++++++++++--
 drivers/video/fbdev/simplefb.c    | 57 ++++++++++++++++++++++---------
 include/linux/fb.h                |  1 +
 5 files changed, 80 insertions(+), 26 deletions(-)
drivers/firmware/sysfb_simplefb.c +1 -1
index b86761904949..179e9d0ef3e9 100644
@@ -99,7 +99,7 @@ __init int sysfb_create_simplefb(const struct screen_info *si,
 
 	/* setup IORESOURCE_MEM as framebuffer memory */
 	memset(&res, 0, sizeof(res));
-	res.flags = IORESOURCE_MEM | IORESOURCE_BUSY;
+	res.flags = IORESOURCE_MEM;
 	res.name = simplefb_resname;
 	res.start = base;
 	res.end = res.start + length - 1;
drivers/gpu/drm/tiny/simpledrm.c +12 -5
index 04146da2d1d8..32550697b1a7 100644
@@ -526,21 +526,28 @@ static int simpledrm_device_init_mm(struct simpledrm_device *sdev)
 {
 	struct drm_device *dev = &sdev->dev;
 	struct platform_device *pdev = sdev->pdev;
-	struct resource *mem;
+	struct resource *res, *mem;
 	void __iomem *screen_base;
 	int ret;
 
-	mem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
-	if (!mem)
+	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	if (!res)
 		return -EINVAL;
 
-	ret = devm_aperture_acquire_from_firmware(dev, mem->start, resource_size(mem));
+	ret = devm_aperture_acquire_from_firmware(dev, res->start, resource_size(res));
 	if (ret) {
 		drm_err(dev, "could not acquire memory range %pr: error %d\n",
-			mem, ret);
+			res, ret);
 		return ret;
 	}
 
+	mem = devm_request_mem_region(&pdev->dev, res->start, resource_size(res),
+				      sdev->dev.driver->name);
+	if (!mem) {
+		drm_err(dev, "could not acquire memory region %pr\n", res);
+		return -EBUSY;
+	}
+
 	screen_base = devm_ioremap_wc(&pdev->dev, mem->start,
 				      resource_size(mem));
 	if (!screen_base)
drivers/video/fbdev/core/fbmem.c +26 -3
index 0fa7ede94fa6..2db31ea7cbb4 100644
@@ -25,6 +25,7 @@
 #include <linux/init.h>
 #include <linux/linux_logo.h>
 #include <linux/proc_fs.h>
+#include <linux/platform_device.h>
 #include <linux/seq_file.h>
 #include <linux/console.h>
 #include <linux/kmod.h>
@@ -1557,18 +1558,36 @@ static void do_remove_conflicting_framebuffers(struct apertures_struct *a,
 	/* check all firmware fbs and kick off if the base addr overlaps */
 	for_each_registered_fb(i) {
 		struct apertures_struct *gen_aper;
+		struct device *dev;
 
 		if (!(registered_fb[i]->flags & FBINFO_MISC_FIRMWARE))
 			continue;
 
 		gen_aper = registered_fb[i]->apertures;
+		dev = registered_fb[i]->device;
 		if (fb_do_apertures_overlap(gen_aper, a) ||
 			(primary && gen_aper && gen_aper->count &&
 			 gen_aper->ranges[0].base == VGA_FB_PHYS)) {
 
 			printk(KERN_INFO "fb%d: switching to %s from %s\n",
 			       i, name, registered_fb[i]->fix.id);
-			do_unregister_framebuffer(registered_fb[i]);
+
+			/*
+			 * If we kick-out a firmware driver, we also want to remove
+			 * the underlying platform device, such as simple-framebuffer,
+			 * VESA, EFI, etc. A native driver will then be able to
+			 * allocate the memory range.
+			 *
+			 * If it's not a platform device, at least print a warning. A
+			 * fix would add code to remove the device from the system.
+			 */
+			if (dev_is_platform(dev)) {
+				registered_fb[i]->forced_out = true;
+				platform_device_unregister(to_platform_device(dev));
+			} else {
+				printk(KERN_WARNING "fb%d: cannot remove device\n", i);
+				do_unregister_framebuffer(registered_fb[i]);
+			}
 		}
 	}
 }
@@ -1898,9 +1917,13 @@ EXPORT_SYMBOL(register_framebuffer);
 void
 unregister_framebuffer(struct fb_info *fb_info)
 {
-	mutex_lock(&registration_lock);
+	bool forced_out = fb_info->forced_out;
+
+	if (!forced_out)
+		mutex_lock(&registration_lock);
 	do_unregister_framebuffer(fb_info);
-	mutex_unlock(&registration_lock);
+	if (!forced_out)
+		mutex_unlock(&registration_lock);
 }
 EXPORT_SYMBOL(unregister_framebuffer);
 
drivers/video/fbdev/simplefb.c +40 -17

103 lines not on this page · raw message

include/linux/fb.h +1 -0
index 3da95842b207..9a14f3f8a329 100644
@@ -502,6 +502,7 @@ struct fb_info {
 	} *apertures;
 
 	bool skip_vt_switch; /* no VT switch on suspend/resume required */
+	bool forced_out; /* set when being removed by another driver */
 };
 
 static inline struct apertures_struct *alloc_apertures(unsigned int max_num) {
Signature · 1 line
2.34.1
#13 lore raw
Quoted 20 lines · Thomas Zimmermann
Quoted 10 lines
Quoted 7 lines

If that works, would you consider protecting pci_request_region() with
   #if not defined(CONFIG_FB_SIMPLE)
   #endif

with a FIXME comment?

Yes, I think that's a good compromise. I'll respin the patch with that.

Before you do that, I have one more patch for you to try. It's all the
changes as before, but now fbdev hot-unplugs the underlying platform device from the device hierarchy. The BOOTFB will not consume parts of
vmwgfx's display memory range any longer. It's now the same behavior as
with simpledrm.

Quoted 1 line

This works for me with simplefb and efifb on i915 hardware.

Yea, that works for me too. Both with simpledrm and simplefb. The patch looks good too.

Do you think you'll be able to get this in stable kernels? If not I'll still need something for vmwgfx to make kernels between 5.15 and whenever this patch gets in boot with fb.

z

#14 lore raw

Hi

Am 20.01.22 um 22:28 schrieb Zack Rusin:

Quoted 21 lines · Thomas Zimmermann
Quoted 20 lines
Quoted 10 lines
Quoted 7 lines

If that works, would you consider protecting pci_request_region() with
   #if not defined(CONFIG_FB_SIMPLE)
   #endif

with a FIXME comment?

Yes, I think that's a good compromise. I'll respin the patch with that.

Before you do that, I have one more patch for you to try. It's all the
changes as before, but now fbdev hot-unplugs the underlying platform device from the device hierarchy. The BOOTFB will not consume parts of
vmwgfx's display memory range any longer. It's now the same behavior as
with simpledrm.

Quoted 8 lines
Quoted 1 line

This works for me with simplefb and efifb on i915 hardware.

Yea, that works for me too. Both with simpledrm and simplefb. The patch looks good too.

Do you think you'll be able to get this in stable kernels? If not I'll still need something for vmwgfx to make kernels between 5.15 and whenever this patch gets in boot with fb.

I'll prepare a patchset with these changes. The actual fix is the change to fbmem.c. This one should go to stable and should be easy to backport. All the other patches are mostly for correctness and can go to drm-misc-next only.

Can I add your Tested-by tag to the patches?

Best regards
Thomas

Quoted 2 lines

z

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#15 lore raw
In reply to Thomas Zimmermann
Quoted 61 lines · Thomas Zimmermann

Hi

Am 19.01.22 um 03:15 schrieb Zack Rusin:

Quoted 40 lines · Javier Martinez Canillas
Quoted 12 lines

Hello Zack,

Quoted 8 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions
are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 5 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY
flag from the memory resource added to the "simple-framebuffer" device
?

I think this is one of those cases where it depends on what we plan to
do after that. Sementically it makes sense to have it in there - the
framebuffer memory is claimed by the "simple-framebuffer" and it's busy, it's just that it creates issues for drivers after unloading. I
think removing it, while making things easier for drivers, would be confusing for people reading the code later, unless there's some kind
of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better solution than having the drm drivers not call pci_request_region[s],
which apart from hyperv and cirrus (iirc bochs does it for resources
other than fb which wouldn't have been claimed by "simple- framebuffer")
is already the case.

Quoted 3 lines

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove pci_request_region[s] from hyperv and cirrus.

I just discussed this a bit with Javier. It's a problem with the simple-framebuffer code, rather then vmwgfx.

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have drivers register/release the range with _BUSY. That would signal the memory belongs to the sysfb device but is not busy unless a driver has
been bound. After simplefb released the range, it should be 'non- busy'
again and available for vmwgfx. Simpledrm does a hot-unplug of the sysfb
device, so the memory range gets released entirely. If you want, I'll prepare some patches for this scenario.

Quoted 5 lines

If this doesn't work, pushing all request/release pairs into drivers would be my next option.

If none of this is feasible, we can still remove pci_request_region() from vmwgfx.

I think that's orthogonal to the fix because having pci_request_region makes vmwgfx behave differently from majority of DRM drivers, e.g. on systems with sysfb enabled with 5.15 vmwgfx fails to boot and leaves the system broken without any fb driver (because while we have *remove_conflicting*_framebuffers we don't have drm_restore_system_fb or such to load back the boot fb after drm driver load fails) but since it's one of the few drivers which does request regions it took a bit for us to notice.

So in this case I'd much rather be like the other drivers rather than correct because it lowers the odds of vmwgfx breaking in the future.

z

#16 lore raw

Hi

Am 19.01.22 um 15:24 schrieb Zack Rusin:

Quoted 78 lines · Thomas Zimmermann
Quoted 61 lines

Hi

Am 19.01.22 um 03:15 schrieb Zack Rusin:

Quoted 40 lines · Javier Martinez Canillas
Quoted 12 lines

Hello Zack,

Quoted 8 lines · Zack Rusin

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions
are held by the platform. In that case remove_conflicting*_framebuffers
only removes the simplefb but not the regions held by sysfb.

Quoted 5 lines

Indeed, that's an issue. I wonder if we should drop the IORESOURCE_BUSY
flag from the memory resource added to the "simple-framebuffer" device
?

I think this is one of those cases where it depends on what we plan to
do after that. Sementically it makes sense to have it in there - the
framebuffer memory is claimed by the "simple-framebuffer" and it's busy, it's just that it creates issues for drivers after unloading. I
think removing it, while making things easier for drivers, would be confusing for people reading the code later, unless there's some kind
of cleanup that would happen with it (e.g. removing IORESOURCE_BUSY altogether and making the drm drivers properly request their resources). At least by itself it doesn't seem to be much better solution than having the drm drivers not call pci_request_region[s],
which apart from hyperv and cirrus (iirc bochs does it for resources
other than fb which wouldn't have been claimed by "simple- framebuffer")
is already the case.

Quoted 3 lines

I do think we should do one of them to make the codebase coherent:
either remove IORESOURCE_BUSY from "simple-framebuffer" or remove pci_request_region[s] from hyperv and cirrus.

I just discussed this a bit with Javier. It's a problem with the simple-framebuffer code, rather then vmwgfx.

IMHO the best solution is to drop IORESOURCE_BUSY from sysfb and have drivers register/release the range with _BUSY. That would signal the memory belongs to the sysfb device but is not busy unless a driver has
been bound. After simplefb released the range, it should be 'non- busy'
again and available for vmwgfx. Simpledrm does a hot-unplug of the sysfb
device, so the memory range gets released entirely. If you want, I'll prepare some patches for this scenario.

Quoted 5 lines

If this doesn't work, pushing all request/release pairs into drivers would be my next option.

If none of this is feasible, we can still remove pci_request_region() from vmwgfx.

I think that's orthogonal to the fix because having pci_request_region makes vmwgfx behave differently from majority of DRM drivers, e.g. on systems with sysfb enabled with 5.15 vmwgfx fails to boot and leaves the system broken without any fb driver (because while we have *remove_conflicting*_framebuffers we don't have drm_restore_system_fb or such to load back the boot fb after drm driver load fails) but since it's one of the few drivers which does request regions it took a bit for us to notice.

Quoted 2 lines

So in this case I'd much rather be like the other drivers rather than correct because it lowers the odds of vmwgfx breaking in the future.

Well, if you want to remove the calls then do so, of course.

I'd rather add the request_memory() calls to all the other drivers that are missing them. Javier suggested to make this an official TODO item. Having the BUSY flag set when a driver is active, still is the correct thing to do.

I'm not sure i understand your comment about drm_restore_system_fb. There is no way of atomically switching drivers and that's always been a problem. Failing at pci_request_memroy() is just one of many possible reasons for the switch to fail. The best you could do is to rearrange the code to do 'remove_conflicting*()' at the latest point possible.

Best regards
Thomas

Quoted 3 lines

z

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#17 lore raw
In reply to Zack Rusin

Hi Zack

Am 17.01.22 um 19:03 schrieb Zack Rusin:

Quoted 5 lines

From: Zack Rusin <[email protected]>

When sysfb_simple is enabled loading vmwgfx fails because the regions are held by the platform. In that case remove_conflicting*_framebuffers only removes the simplefb but not the regions held by sysfb.

I don't understand this sentence. There's only one memory resource claimed by the sysfb code. What else would block vmwgfx?

I appears to me as if simplefb should release the region (or the simple-framebuffer device).

simpledrm does a hot-unplug of the simple-framebuffer, so the region should be released afterwards.

Best regard
Thomas

Quoted 29 lines

Like the other drm drivers we need to stop requesting all the pci regions to let the driver load with platform code enabled.
This allows vmwgfx to load correctly on systems with sysfb_simple enabled.

Fixes 523375c943e5 ("drm/vmwgfx: Port vmwgfx to arm64")
---
  drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 4 ----
  1 file changed, 4 deletions(-)
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +0 -4
index fe36efdb7ff5..27feb19f3324 100644
@@ -724,10 +724,6 @@ static int vmw_setup_pci_resources(struct vmw_private *dev,
 
  	pci_set_master(pdev);
 
-	ret = pci_request_regions(pdev, "vmwgfx probe");
-	if (ret)
-		return ret;
-
  	dev->pci_id = pci_id;
  	if (pci_id == VMWGFX_PCI_ID_SVGA3) {
  		rmmio_start = pci_resource_start(pdev, 0);
Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev