BGRT doesn't work for me on efi-next

Hello Sai and others,

The change to use early_mem*() instead of early_io*() in 50a0cb56 does not work on my machine. Last week I discussed some BGRT changes and I created a patch for that, but can't test it on efi-next because of this.

I get this (when booting 50a0cb56, without any of my changes):

[    0.026936] ------------[ cut here ]------------
[    0.026941] WARNING: CPU: 0 PID: 0 at mm/early_ioremap.c:137
__early_ioremap+0x102/0x1bb()
[    0.026941] Modules linked in:
[    0.026944] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.4.0-rc1 #2
[    0.026945] Hardware name: Dell Inc. XPS 13 9343/09K8G1, BIOS A05
07/14/2015
[    0.026946]  0000000000000000 900f03d5a116524d ffffffff81c03e60
ffffffff813a3fff
[    0.026948]  0000000000000000 ffffffff81c03e98 ffffffff810a0852
00000000d7b76000
[    0.026949]  0000000000000000 0000000000000001 0000000000000001
000000000000017c
[    0.026951] Call Trace:
[    0.026955]  [<ffffffff813a3fff>] dump_stack+0x44/0x55
[    0.026958]  [<ffffffff810a0852>] warn_slowpath_common+0x82/0xc0
[    0.026959]  [<ffffffff810a099a>] warn_slowpath_null+0x1a/0x20
[    0.026961]  [<ffffffff81d8c395>] __early_ioremap+0x102/0x1bb
[    0.026962]  [<ffffffff81d8c602>] early_memremap+0x13/0x15
[    0.026964]  [<ffffffff81d78361>] efi_bgrt_init+0x162/0x1ad
[    0.026966]  [<ffffffff81d778ec>] efi_late_init+0x9/0xb
[    0.026968]  [<ffffffff81d58ff5>] start_kernel+0x46f/0x49f
[    0.026970]  [<ffffffff81d58120>] ? early_idt_handler_array+0x120/0x120
[    0.026972]  [<ffffffff81d58339>] x86_64_start_reservations+0x2a/0x2c
[    0.026974]  [<ffffffff81d58485>] x86_64_start_kernel+0x14a/0x16d
[    0.026977] ---[ end trace f9b3812eb8e24c58 ]---
[    0.026978] efi_bgrt: Ignoring BGRT: failed to map image memory

This is the second early_memremap call in efi_bgrt_init triggering WARN_ON(nrpages > NR_FIX_BTMAPS). Can you comment on this?

(Pulling in Ingo and Juergen because of commit 954e12f7a800)

Quoted 35 lines · Môshe van der Sterre

Hello Sai and others,

The change to use early_mem*() instead of early_io*() in 50a0cb56 does not work on my machine. Last week I discussed some BGRT changes and I created a patch for that, but can't test it on efi-next because of this.

I get this (when booting 50a0cb56, without any of my changes):

[    0.026936] ------------[ cut here ]------------
[    0.026941] WARNING: CPU: 0 PID: 0 at mm/early_ioremap.c:137
__early_ioremap+0x102/0x1bb()
[    0.026941] Modules linked in:
[    0.026944] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.4.0-rc1 #2
[    0.026945] Hardware name: Dell Inc. XPS 13 9343/09K8G1, BIOS A05
07/14/2015
[    0.026946]  0000000000000000 900f03d5a116524d ffffffff81c03e60
ffffffff813a3fff
[    0.026948]  0000000000000000 ffffffff81c03e98 ffffffff810a0852
00000000d7b76000
[    0.026949]  0000000000000000 0000000000000001 0000000000000001
000000000000017c
[    0.026951] Call Trace:
[    0.026955]  [<ffffffff813a3fff>] dump_stack+0x44/0x55
[    0.026958]  [<ffffffff810a0852>] warn_slowpath_common+0x82/0xc0
[    0.026959]  [<ffffffff810a099a>] warn_slowpath_null+0x1a/0x20
[    0.026961]  [<ffffffff81d8c395>] __early_ioremap+0x102/0x1bb
[    0.026962]  [<ffffffff81d8c602>] early_memremap+0x13/0x15
[    0.026964]  [<ffffffff81d78361>] efi_bgrt_init+0x162/0x1ad
[    0.026966]  [<ffffffff81d778ec>] efi_late_init+0x9/0xb
[    0.026968]  [<ffffffff81d58ff5>] start_kernel+0x46f/0x49f
[    0.026970]  [<ffffffff81d58120>] ? early_idt_handler_array+0x120/0x120
[    0.026972]  [<ffffffff81d58339>] x86_64_start_reservations+0x2a/0x2c
[    0.026974]  [<ffffffff81d58485>] x86_64_start_kernel+0x14a/0x16d
[    0.026977] ---[ end trace f9b3812eb8e24c58 ]---
[    0.026978] efi_bgrt: Ignoring BGRT: failed to map image memory

Quoted 2 lines

This is the second early_memremap call in efi_bgrt_init triggering WARN_ON(nrpages > NR_FIX_BTMAPS). Can you comment on this?

Hmm... yeah NR_FIX_BTMAPS == 64, so early_memremap() is limited to mapping ~200KB at once. That's not very big, your BGRT data is likely much larger than that.

Obviously we can't use efi_lookup_mapped_addr() any more, so it makes sense to come up with a much robust way to memremap the BGRT image.

The immediate solution that comes to mind is using memremap() instead of the early_* version, since the late version won't use the FIXMAP area and will allow us to map a much larger region into the kernel virtual address space.

Digging through the history it appears I was the one who made the switch from ioremap() to early_memremap() in commit 081cd62a010f ("x86/efi: Allow mapping BGRT on x86-32"), which I suspect was because a generic memremap() implementation didn't exist at the time. Dan Williams introduced one in Aug 2015 with commit 92281dee825f ("arch:
introduce memremap()").

Somewhat surprisingly, Juergen switched the BGRT driver from early_memremap() to early_ioremap() in commit 954e12f7a800 ("x86/mm,
efi: Use early_ioremap() in arch/x86/platform/efi/efi-bgrt.c") although I can't figure out why because this isn't I/O memory. And now we're attempting to switch it back.

Juergen, could you provide a rationale for commit 954e12f7a800? All the commit message says is that it's "an I/O-area", but that isn't true.

In the meantime Môshe, could you try this patch ontop of the EFI 'next' branch? (Note it may not work/compile, but you get the gist)

---

arch/x86/platform/efi/efi-bgrt.c +4 -4

index bf51f4c02562..b0970661870a 100644
@@ -72,14 +72,14 @@ void __init efi_bgrt_init(void)
 		return;
 	}
 
-	image = early_memremap(bgrt_tab->image_address, sizeof(bmp_header));
+	image = memremap(bgrt_tab->image_address, sizeof(bmp_header), MEMREMAP_WB);
 	if (!image) {
 		pr_err("Ignoring BGRT: failed to map image header memory\n");
 		return;
 	}
 
 	memcpy(&bmp_header, image, sizeof(bmp_header));
-	early_memunmap(image, sizeof(bmp_header));
+	memunmap(image);
 	bgrt_image_size = bmp_header.size;
 
 	bgrt_image = kmalloc(bgrt_image_size, GFP_KERNEL | __GFP_NOWARN);

@@ -89,7 +89,7 @@ void __init efi_bgrt_init(void)
 		return;
 	}
 
-	image = early_memremap(bgrt_tab->image_address, bmp_header.size);
+	image = memremap(bgrt_tab->image_address, bmp_header.size, MEMREMAP_WB);
 	if (!image) {
 		pr_err("Ignoring BGRT: failed to map image memory\n");
 		kfree(bgrt_image);

@@ -98,5 +98,5 @@ void __init efi_bgrt_init(void)
 	}
 
 	memcpy(bgrt_image, image, bgrt_image_size);
-	early_memunmap(image, bmp_header.size);
+	memunmap(image);
 }

(Pulling in Ingo and Juergen because of commit 954e12f7a800)

Quoted 35 lines · Môshe van der Sterre

Hello Sai and others,

The change to use early_mem*() instead of early_io*() in 50a0cb56 does not work on my machine. Last week I discussed some BGRT changes and I created a patch for that, but can't test it on efi-next because of this.

I get this (when booting 50a0cb56, without any of my changes):

[    0.026936] ------------[ cut here ]------------
[    0.026941] WARNING: CPU: 0 PID: 0 at mm/early_ioremap.c:137
__early_ioremap+0x102/0x1bb()
[    0.026941] Modules linked in:
[    0.026944] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.4.0-rc1 #2
[    0.026945] Hardware name: Dell Inc. XPS 13 9343/09K8G1, BIOS A05
07/14/2015
[    0.026946]  0000000000000000 900f03d5a116524d ffffffff81c03e60
ffffffff813a3fff
[    0.026948]  0000000000000000 ffffffff81c03e98 ffffffff810a0852
00000000d7b76000
[    0.026949]  0000000000000000 0000000000000001 0000000000000001
000000000000017c
[    0.026951] Call Trace:
[    0.026955]  [<ffffffff813a3fff>] dump_stack+0x44/0x55
[    0.026958]  [<ffffffff810a0852>] warn_slowpath_common+0x82/0xc0
[    0.026959]  [<ffffffff810a099a>] warn_slowpath_null+0x1a/0x20
[    0.026961]  [<ffffffff81d8c395>] __early_ioremap+0x102/0x1bb
[    0.026962]  [<ffffffff81d8c602>] early_memremap+0x13/0x15
[    0.026964]  [<ffffffff81d78361>] efi_bgrt_init+0x162/0x1ad
[    0.026966]  [<ffffffff81d778ec>] efi_late_init+0x9/0xb
[    0.026968]  [<ffffffff81d58ff5>] start_kernel+0x46f/0x49f
[    0.026970]  [<ffffffff81d58120>] ? early_idt_handler_array+0x120/0x120
[    0.026972]  [<ffffffff81d58339>] x86_64_start_reservations+0x2a/0x2c
[    0.026974]  [<ffffffff81d58485>] x86_64_start_kernel+0x14a/0x16d
[    0.026977] ---[ end trace f9b3812eb8e24c58 ]---
[    0.026978] efi_bgrt: Ignoring BGRT: failed to map image memory

Quoted 2 lines

This is the second early_memremap call in efi_bgrt_init triggering WARN_ON(nrpages > NR_FIX_BTMAPS). Can you comment on this?

Hmm... yeah NR_FIX_BTMAPS == 64, so early_memremap() is limited to mapping ~200KB at once. That's not very big, your BGRT data is likely much larger than that.

Obviously we can't use efi_lookup_mapped_addr() any more, so it makes sense to come up with a much robust way to memremap the BGRT image.

The immediate solution that comes to mind is using memremap() instead of the early_* version, since the late version won't use the FIXMAP area and will allow us to map a much larger region into the kernel virtual address space.

Digging through the history it appears I was the one who made the switch from ioremap() to early_memremap() in commit 081cd62a010f ("x86/efi: Allow mapping BGRT on x86-32"), which I suspect was because a generic memremap() implementation didn't exist at the time. Dan Williams introduced one in Aug 2015 with commit 92281dee825f ("arch:
introduce memremap()").

Somewhat surprisingly, Juergen switched the BGRT driver from early_memremap() to early_ioremap() in commit 954e12f7a800 ("x86/mm,
efi: Use early_ioremap() in arch/x86/platform/efi/efi-bgrt.c") although I can't figure out why because this isn't I/O memory. And now we're attempting to switch it back.

Juergen, could you provide a rationale for commit 954e12f7a800? All the commit message says is that it's "an I/O-area", but that isn't true.

In the meantime Môshe, could you try this patch ontop of the EFI 'next' branch? (Note it may not work/compile, but you get the gist)

---

arch/x86/platform/efi/efi-bgrt.c +4 -5

index bf51f4c02562..b0970661870a 100644
@@ -72,14 +72,14 @@ void __init efi_bgrt_init(void)
  		return;
  	}
 
-	image = early_memremap(bgrt_tab->image_address, sizeof(bmp_header));
+	image = memremap(bgrt_tab->image_address, sizeof(bmp_header), MEMREMAP_WB);
  	if (!image) {
  		pr_err("Ignoring BGRT: failed to map image header memory\n");
  		return;
  	}
 
  	memcpy(&bmp_header, image, sizeof(bmp_header));
-	early_memunmap(image, sizeof(bmp_header));
+	memunmap(image);
  	bgrt_image_size = bmp_header.size;
 
  	bgrt_image = kmalloc(bgrt_image_size, GFP_KERNEL | __GFP_NOWARN);

@@ -89,7 +89,7 @@ void __init efi_bgrt_init(void)
  		return;
  	}
 
-	image = early_memremap(bgrt_tab->image_address, bmp_header.size);
+	image = memremap(bgrt_tab->image_address, bmp_header.size, MEMREMAP_WB);
  	if (!image) {
  		pr_err("Ignoring BGRT: failed to map image memory\n");
  		kfree(bgrt_image);

@@ -98,5 +98,5 @@ void __init efi_bgrt_init(void)
  	}
 
  	memcpy(bgrt_image, image, bgrt_image_size);
-	early_memunmap(image, bmp_header.size);
+	memunmap(image);
  }
--

To unsubscribe from this list: send the line "unsubscribe linux-efi" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html