Re: [PATCH] pci/vgaarb: make vga_is_firmware_default() arch independent
From: Sui Jingfeng
Date: Wed May 31 2023 - 05:48:39 EST
Hi,
CCing loogson EFI co-workers:
On 2023/5/31 04:17, Bjorn Helgaas wrote:
On Mon, May 29, 2023 at 11:45:04PM +0800, Sui Jingfeng wrote:
The vga_is_firmware_default() function will work on non-x86 architectures
as long as the arch has UEFI GOP support, which passes the firmware
framebuffer base address and size.
This patch makes the vga_is_firmware_default() function arch-independent.
This could help the VGAARB subsystem make the right choice for multiple
GPU systems. Usually an integrated one and a discrete one for desktop
computers. Depending on the firmware framebuffer being put into which
GPU's VRAM, VGAARB could inherit the firmware's choice, which in turn,
is the exact choice of the user.
Is there a system that needs this change? If so, the commit log
should mention it.
We omit the architectures-specific//deliberately in the commit message.
When I create this patch, I'm also wondering whether we should enumerate
the usage cases.
But as far as I can see it, vgaarb is more about the PCI peripherals.
It's more about the video adapter, firmware framebuffer, etc.
It seems that there is no good reason why this function should be
monopolized by X86 and IA64.
Loongson LoongArch is one of the architectures with EFIFB support, which
could utilize this function to implement something fun.
But it's OK, we could improve the commit message in the next version of
this patch.
It's definitely nice to remove #ifdefs, but it's better if we have an
actual reason and some testing of another arch that makes use of this.
As far as I know, Arm and Arm64 also have EFIFB support.
while this patch is not relevant to firmware (efi, uboot, pmon, etc.).
It is only relevant to the global screen_info.
As long as the firmware passes screen_info to the kernel.
On one of my LoongArch machines with AMDGPU mounted,
The UEFI firmware puts its firmware framebuffer at the amdgpu's vram BAR.
With this patch applied, it works as expected. see below:
AMD RX550:
[ 0.358155] pci 0000:03:00.0: BAR 0: assigned [mem
0xe0020000000-0xe002fffffff 64bit pref]
[ 0.358165] pci 0000:03:00.0: BAR 2: assigned [mem
0xe0030000000-0xe00301fffff 64bit pref]
[ 0.358175] pci 0000:03:00.0: BAR 5: assigned [mem
0xe0039200000-0xe003923ffff]
[ 0.358180] pci 0000:03:00.0: BAR 6: assigned [mem
0xe0039240000-0xe003925ffff pref]
[ 0.358182] pci 0000:03:00.1: BAR 0: assigned [mem
0xe0039260000-0xe0039263fff 64bit]
[ 0.358192] pci 0000:03:00.0: BAR 4: assigned [io 0x5000-0x50ff]
Loongson integrated graphics:
[ 0.355740] pci 0000:00:06.0: [0014:7a15] type 00 class 0x040000
[ 0.355747] pci 0000:00:06.0: reg 0x10: [mem
0xe0035300000-0xe003533ffff 64bit]
[ 0.355753] pci 0000:00:06.0: reg 0x18: [mem
0xe00353c6000-0xe00353c6fff 64bit]
[ 0.355758] pci 0000:00:06.0: reg 0x20: [mem
0xe0035360000-0xe003536ffff 64bit]
[ 0.355797] pci 0000:00:06.1: [0014:7a06] type 00 class 0x030000
[ 0.355805] pci 0000:00:06.1: reg 0x10: [mem
0xe0035370000-0xe003537ffff 64bit]
VGAARB:
[ 0.362180] pci 0000:00:06.1: vgaarb: base: 0xe0020000000 size: 0xf3c00
[ 0.362184] pci 0000:00:06.1: vgaarb: setting as boot VGA device
[ 0.362186] pci 0000:00:06.1: vgaarb: bridge control possible
[ 0.362188] pci 0000:00:06.1: vgaarb: VGA device added:
decodes=io+mem,owns=io+mem,locks=none
[ 0.362199] pci 0000:03:00.0: vgaarb: base: 0xe0020000000 size: 0xf3c00
[ 0.362202] pci 0000:03:00.0: vgaarb: is firmware default
[ 0.362204] pci 0000:03:00.0: vgaarb: setting as boot VGA device
(overriding previous)
[ 0.362206] pci 0000:03:00.0: vgaarb: bridge control possible
[ 0.362208] pci 0000:03:00.0: vgaarb: VGA device added:
decodes=io+mem,owns=none,locks=none
[ 0.362211] vgaarb: loaded
[ 16.758480] loongson 0000:00:06.1: vgaarb: changed VGA decodes:
olddecodes=io+mem,decodes=none:owns=io+mem
[ 16.768075] amdgpu 0000:03:00.0: vgaarb: changed VGA decodes:
olddecodes=io+mem,decodes=none:owns=none
The above log is extracted from the dmesg log;
This log can prove that the discrete GPU finally overrides the
integrated one.
In this case, the discrete GPU will be the default boot device.
This meets normal users' expectations.
But what if we want to use the integrated one as the default boot device
and have it opened by the X server as the primary GPU by default?
Without opening the computer's machine box, unmount the discrete GPU
every time.
On a multi-GPU system, this function gives us a chance to mandate which
card to use by default.
There is no need to invent a complex parameter-passing strategy between
the firmware and the kernel.
The UEFI firmware could simply place the firmware framebuffer at the
desired GPU's vram bar.
vgaarb could decode the user's setting by comparing address ranges.
But still, this is more about PCI peripherals and/or firmware framebuffers.
An arch that can't use this machinism should jump in here and explain
why their arch can't support it. Am I right?
This is also the reason I'm CCing my co-workers who make the Loongson
UEFI firmware,
There are early-stage UEFI firmwares for Loongson machines released,
which either put the EFIFB in the wrong bar or don't have UEFI GOP
support at all.
Which causes the kernel space (GPU/DC) device driver to have to pay for it.
no more please. I hope they can do the right thing.
Also, take a look at the git history and match the subject line and
commit log style (prefix, capitalization, imperative voice).
Ok, nice comments.
I'm not native English speaker, sometimes I can't sense the subtle
differences of similar word.
But I will improve it at next version, please give more tolerance, thanks.
Signed-off-by: Sui Jingfeng <suijingfeng@xxxxxxxxxxx>
---
drivers/pci/vgaarb.c | 7 ++-----
1 file changed, 2 insertions(+), 5 deletions(-)
diff --git a/drivers/pci/vgaarb.c b/drivers/pci/vgaarb.c
index 5a696078b382..f81b6c54e327 100644
--- a/drivers/pci/vgaarb.c
+++ b/drivers/pci/vgaarb.c
@@ -61,7 +61,6 @@ static bool vga_arbiter_used;
static DEFINE_SPINLOCK(vga_lock);
static DECLARE_WAIT_QUEUE_HEAD(vga_wait_queue);
-
static const char *vga_iostate_to_str(unsigned int iostate)
{
/* Ignore VGA_RSRC_IO and VGA_RSRC_MEM */
@@ -545,7 +544,6 @@ EXPORT_SYMBOL(vga_put);
static bool vga_is_firmware_default(struct pci_dev *pdev)
{
-#if defined(CONFIG_X86) || defined(CONFIG_IA64)
u64 base = screen_info.lfb_base;
u64 size = screen_info.lfb_size;
struct resource *r;
@@ -571,7 +569,7 @@ static bool vga_is_firmware_default(struct pci_dev *pdev)
return true;
}
-#endif
+
return false;
}
@@ -865,8 +863,7 @@ static bool vga_arbiter_del_pci_device(struct pci_dev *pdev)
}
/* this is called with the lock */
-static inline void vga_update_device_decodes(struct vga_device *vgadev,
- int new_decodes)
+static void vga_update_device_decodes(struct vga_device *vgadev, int new_decodes)
I don't mind removing the "inline" here, but it shouldn't be combined
with the rest of the patch. When it's combined, I can't tell whether
there's a reason we need this change or if it's just a cleanup.
Previously, make a single patch is for convenient.
OK, I split it into two patch at the next version.
{
struct device *dev = &vgadev->pdev->dev;
int old_decodes, decodes_removed, decodes_unlocked;
--
2.25.1
--
Jingfeng