Re: [PATCH] fbdev: efifb: do not load efifb if PCI BAR has changed but not fixuped

[Shixiong Ou]

在 2025/7/7 17:28, Thomas Zimmermann 写道:
Hi

Am 07.07.25 um 11:24 schrieb Shixiong Ou:

在 2025/6/27 17:13, Thomas Zimmermann 写道:
Hi

Am 26.06.25 um 11:49 schrieb oushixiong1...@163.com:
From: Shixiong Ou <oushixi...@kylinos.cn>

[WHY]
On an ARM machine, the following log is present:
[    0.900884] efifb: framebuffer at 0x1020000000, using 3072k, 
total 3072k
[    2.297884] amdgpu 0000:04:00.0: 
remove_conflicting_pci_framebuffers: bar 0: 0x1000000000 -> 
0x100fffffff
[    2.297886] amdgpu 0000:04:00.0: 
remove_conflicting_pci_framebuffers: bar 2: 0x1010000000 -> 
0x10101fffff
[    2.297888] amdgpu 0000:04:00.0: 
remove_conflicting_pci_framebuffers: bar 5: 0x58200000 -> 0x5823ffff

It show that the efifb framebuffer base is out of PCI BAR, and this
results in both efi-framebuffer and amdgpudrmfb co-existing.

The fbcon will be bound to efi-framebuffer by default and cannot be 
used.

[HOW]
Do not load efifb driver if PCI BAR has changed but not fixuped.
In the following cases:
    1. screen_info_lfb_pdev is NULL.
    2. __screen_info_relocation_is_valid return false.

Apart from ruling out invalid screen_info, did you figure out why 
the relocation tracking didn't work? It would be good to fix this if 
possible.

Best regards
Thomas

I haven’t figure out the root cause yet.

This issue is quite rare and might be related to the EFI firmware.
However, I wonder if we could add some handling when no PCI resources 
are found in screen_info_fixup_lfb(), as a temporary workaround for 
the problem I mentioned earlier.

As I said elsewhere in the thread, you can clear the screen_info's 
video type in the branch at [1] to disable it entirely. We should have 
probably done this anyway. Knowing the cause of the issue would still 
be nice though.

Best regards
Thomas

[1] 
https://elixir.bootlin.com/linux/v6.15.5/source/drivers/video/screen_info_pci.c#L44 


thanks for you suggestion, while there are two cases:
    1. screen_info_lfb_pdev is NULL.
    2. __screen_info_relocation_is_valid return false.

should we do it at [1] too?

Best regards
Shixiong Ou

[1] 
https://elixir.bootlin.com/linux/v6.15.5/source/drivers/video/screen_info_pci.c#L47


Best regards
Shixiong Ou


Signed-off-by: Shixiong Ou <oushixi...@kylinos.cn>
---
  drivers/video/fbdev/efifb.c     |  4 ++++
  drivers/video/screen_info_pci.c | 24 ++++++++++++++++++++++++
  include/linux/screen_info.h     |  5 +++++
  3 files changed, 33 insertions(+)

diff --git a/drivers/video/fbdev/efifb.c b/drivers/video/fbdev/efifb.c
index 0e1bd3dba255..de8d016c9a66 100644
--- a/drivers/video/fbdev/efifb.c
+++ b/drivers/video/fbdev/efifb.c
@@ -303,6 +303,10 @@ static void efifb_setup(struct screen_info 
*si, char *options)
    static inline bool fb_base_is_valid(struct screen_info *si)
  {
+    /* check whether fb_base has changed but not fixuped */
+    if (!screen_info_is_useful())
+        return false;
+
      if (si->lfb_base)
          return true;
  diff --git a/drivers/video/screen_info_pci.c 
b/drivers/video/screen_info_pci.c
index 66bfc1d0a6dc..ac57dcaf0cac 100644
--- a/drivers/video/screen_info_pci.c
+++ b/drivers/video/screen_info_pci.c
@@ -9,6 +9,8 @@ static struct pci_dev *screen_info_lfb_pdev;
  static size_t screen_info_lfb_bar;
  static resource_size_t screen_info_lfb_res_start; // original 
start of resource
  static resource_size_t screen_info_lfb_offset; // framebuffer 
offset within resource
+static bool screen_info_changed;
+static bool screen_info_fixuped;
    static bool __screen_info_relocation_is_valid(const struct 
screen_info *si, struct resource *pr)
  {
@@ -24,6 +26,24 @@ static bool 
__screen_info_relocation_is_valid(const struct screen_info *si, stru
      return true;
  }
  +bool screen_info_is_useful(void)
+{
+    unsigned int type;
+    const struct screen_info *si = &screen_info;
+
+    type = screen_info_video_type(si);
+    if (type != VIDEO_TYPE_EFI)
+        return true;
+
+    if (screen_info_changed && !screen_info_fixuped) {
+        pr_warn("The screen_info has changed but not fixuped");
+        return false;
+    }
+
+    pr_info("The screen_info is useful");
+    return true;
+}
+
  void screen_info_apply_fixups(void)
  {
      struct screen_info *si = &screen_info;
@@ -32,18 +52,22 @@ void screen_info_apply_fixups(void)
          struct resource *pr = 
&screen_info_lfb_pdev->resource[screen_info_lfb_bar];
            if (pr->start != screen_info_lfb_res_start) {
+            screen_info_changed = true;
              if (__screen_info_relocation_is_valid(si, pr)) {
                  /*
                   * Only update base if we have an actual
                   * relocation to a valid I/O range.
                   */
                  __screen_info_set_lfb_base(si, pr->start + 
screen_info_lfb_offset);
+                screen_info_fixuped = true;
                  pr_info("Relocating firmware framebuffer to 
offset %pa[d] within %pr\n",
                      &screen_info_lfb_offset, pr);
              } else {
                  pr_warn("Invalid relocating, disabling firmware 
framebuffer\n");
              }
          }
+    } else {
+        screen_info_changed = true;
      }
  }
  diff --git a/include/linux/screen_info.h 
b/include/linux/screen_info.h
index 923d68e07679..632cdbb1adbe 100644
--- a/include/linux/screen_info.h
+++ b/include/linux/screen_info.h
@@ -138,9 +138,14 @@ ssize_t screen_info_resources(const struct 
screen_info *si, struct resource *r,
  u32 __screen_info_lfb_bits_per_pixel(const struct screen_info *si);
    #if defined(CONFIG_PCI)
+bool screen_info_is_useful(void);
  void screen_info_apply_fixups(void);
  struct pci_dev *screen_info_pci_dev(const struct screen_info *si);
  #else
+bool screen_info_is_useful(void)
+{
+    return true;
+}
  static inline void screen_info_apply_fixups(void)
  { }
  static inline struct pci_dev *screen_info_pci_dev(const struct 
screen_info *si)