Skip to contentSkip to conversation

x86/efi: Do not release sub-1MB memory regions when the crashkernel option is specified

PATCH In review Lianbo Jiang 3 participants 10 messages Apr 7, 2021 – May 24, 2021 3 messages not in the archive

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/efi: Do not release sub-1MB memory regions when the crashkernel option is specified

Some sub-1MB memory regions may be reserved by EFI boot services, and the memory regions will be released later in the efi_free_boot_services().

Currently, always reserve all sub-1MB memory regions when the crashkernel option is specified, but unfortunately EFI boot services may have already reserved some sub-1MB memory regions before the crash_reserve_low_1M() is called, which makes that the crash_reserve_low_1M() only own the remaining sub-1MB memory regions, not all sub-1MB memory regions, because, subsequently EFI boot services will free its own sub-1MB memory regions. Eventually, DMA will be able to allocate memory from the sub-1MB area and cause the following error:

crash> kmem -s |grep invalid
kmem: dma-kmalloc-512: slab: ffffd52c40001900 invalid freepointer: ffff9403c0067300
kmem: dma-kmalloc-512: slab: ffffd52c40001900 invalid freepointer: ffff9403c0067300 crash> vtop ffff9403c0067300
VIRTUAL PHYSICAL
ffff9403c0067300 67300 --->The physical address falls into this range [0x0000000000063000-0x000000000008efff]

kernel debugging log:
...

[    0.008927] memblock_reserve: [0x0000000000010000-0x0000000000013fff] efi_reserve_boot_services+0x85/0xd0
[    0.008930] memblock_reserve: [0x0000000000063000-0x000000000008efff] efi_reserve_boot_services+0x85/0xd0
...
[    0.009425] memblock_reserve: [0x0000000000000000-0x00000000000fffff] crash_reserve_low_1M+0x2c/0x49
...
[    0.010586] Zone ranges:
[    0.010587]   DMA      [mem 0x0000000000001000-0x0000000000ffffff]
[    0.010589]   DMA32    [mem 0x0000000001000000-0x00000000ffffffff]
[    0.010591]   Normal   [mem 0x0000000100000000-0x0000000c7fffffff]
[    0.010593]   Device   empty
...
[    8.814894] __memblock_free_late: [0x0000000000063000-0x000000000008efff] efi_free_boot_services+0x14b/0x23b
[    8.815793] __memblock_free_late: [0x0000000000010000-0x0000000000013fff] efi_free_boot_services+0x14b/0x23b

Do not release sub-1MB memory regions even though they are reserved by EFI boot services, so that always reserve all sub-1MB memory regions when the crashkernel option is specified.

---
 arch/x86/platform/efi/quirks.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
arch/x86/platform/efi/quirks.c +14 -0
index 67d93a243c35..637f932c4fd4 100644
@@ -18,6 +18,7 @@
 #include <asm/cpu_device_id.h>
 #include <asm/realmode.h>
 #include <asm/reboot.h>
+#include <asm/cmdline.h>
 
 #define EFI_MIN_RESERVE 5120
 
@@ -303,6 +304,19 @@ void __init efi_arch_mem_reserve(phys_addr_t addr, u64 size)
  */
 static __init bool can_free_region(u64 start, u64 size)
 {
+	/*
+	 * Some sub-1MB memory regions may be reserved by EFI boot
+	 * services, and these memory regions will be released later
+	 * in the efi_free_boot_services().
+	 *
+	 * Do not release sub-1MB memory regions even though they are
+	 * reserved by EFI boot services, because, always reserve all
+	 * sub-1MB memory when the crashkernel option is specified.
+	 */
+	if (cmdline_find_option(boot_command_line, "crashkernel", NULL, 0) > 0
+		&& (start + size < (1<<20)))
+		return false;
+
 	if (start + size > __pa_symbol(_text) && start <= __pa_symbol(_end))
 		return false;
 
Signature · 1 line
2.17.1
#1 lore raw
Quoted 12 lines · Lianbo Jiang

Some sub-1MB memory regions may be reserved by EFI boot services, and the memory regions will be released later in the efi_free_boot_services().

Currently, always reserve all sub-1MB memory regions when the crashkernel option is specified, but unfortunately EFI boot services may have already reserved some sub-1MB memory regions before the crash_reserve_low_1M() is called, which makes that the crash_reserve_low_1M() only own the remaining sub-1MB memory regions, not all sub-1MB memory regions, because, subsequently EFI boot services will free its own sub-1MB memory regions. Eventually, DMA will be able to allocate memory from the sub-1MB area and cause the following error:

So this patch is fixing a problem found in crash utility. We ever met the similar issue, later fixed by always reserving low 1M in commit 6f599d84231fd27 ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified"). Seems the commit is not fixing it completely.

Quoted 9 lines

crash> kmem -s |grep invalid
kmem: dma-kmalloc-512: slab: ffffd52c40001900 invalid freepointer: ffff9403c0067300
kmem: dma-kmalloc-512: slab: ffffd52c40001900 invalid freepointer: ffff9403c0067300 crash> vtop ffff9403c0067300
VIRTUAL PHYSICAL
ffff9403c0067300 67300 --->The physical address falls into this range [0x0000000000063000-0x000000000008efff]

kernel debugging log:
...

Quoted 13 lines
[    0.008927] memblock_reserve: [0x0000000000010000-0x0000000000013fff] efi_reserve_boot_services+0x85/0xd0
[    0.008930] memblock_reserve: [0x0000000000063000-0x000000000008efff] efi_reserve_boot_services+0x85/0xd0
...
[    0.009425] memblock_reserve: [0x0000000000000000-0x00000000000fffff] crash_reserve_low_1M+0x2c/0x49
...
[    0.010586] Zone ranges:
[    0.010587]   DMA      [mem 0x0000000000001000-0x0000000000ffffff]
[    0.010589]   DMA32    [mem 0x0000000001000000-0x00000000ffffffff]
[    0.010591]   Normal   [mem 0x0000000100000000-0x0000000c7fffffff]
[    0.010593]   Device   empty
...
[    8.814894] __memblock_free_late: [0x0000000000063000-0x000000000008efff] efi_free_boot_services+0x14b/0x23b
[    8.815793] __memblock_free_late: [0x0000000000010000-0x0000000000013fff] efi_free_boot_services+0x14b/0x23b

In commit 6f599d84231fd27, we call crash_reserve_low_1M() to lock the whole low 1M area if crashkernel is specified in kernel cmdline. But earlier efi_reserve_boot_services() invokation will break the intention of the whole low 1M reserving. In efi_reserve_boot_services(), if any memory under low 1M hasn't been reserved, it will call memblock_reserve() to reserve it and leave it to efi_free_boot_services() to free.

Hi Lianbo,

Please correct me if I am wrong or anything is missed. IIUC, can we move efi_reserve_boot_services() after reserve_real_mode() to fix this bug? Or move reserve_real_mode() before efi_reserve_boot_services() since those real mode regions are all under 1M? Assume efi boot code/data won't rely on low 1M area any more at this moment.

Thanks
Baoquan

Quoted 45 lines

Do not release sub-1MB memory regions even though they are reserved by EFI boot services, so that always reserve all sub-1MB memory regions when the crashkernel option is specified.

---
 arch/x86/platform/efi/quirks.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
arch/x86/platform/efi/quirks.c +14 -0
index 67d93a243c35..637f932c4fd4 100644
@@ -18,6 +18,7 @@
 #include <asm/cpu_device_id.h>
 #include <asm/realmode.h>
 #include <asm/reboot.h>
+#include <asm/cmdline.h>
 
 #define EFI_MIN_RESERVE 5120
 
@@ -303,6 +304,19 @@ void __init efi_arch_mem_reserve(phys_addr_t addr, u64 size)
  */
 static __init bool can_free_region(u64 start, u64 size)
 {
+	/*
+	 * Some sub-1MB memory regions may be reserved by EFI boot
+	 * services, and these memory regions will be released later
+	 * in the efi_free_boot_services().
+	 *
+	 * Do not release sub-1MB memory regions even though they are
+	 * reserved by EFI boot services, because, always reserve all
+	 * sub-1MB memory when the crashkernel option is specified.
+	 */
+	if (cmdline_find_option(boot_command_line, "crashkernel", NULL, 0) > 0
+		&& (start + size < (1<<20)))
+		return false;
+
 	if (start + size > __pa_symbol(_text) && start <= __pa_symbol(_end))
 		return false;
 
Signature · 1 line
2.17.1
author
#2 lore raw

Hi, Baoquan
Thank you for the comment.

Quoted 14 lines · Baoquan He
Quoted 12 lines · Lianbo Jiang

Some sub-1MB memory regions may be reserved by EFI boot services, and the memory regions will be released later in the efi_free_boot_services().

Currently, always reserve all sub-1MB memory regions when the crashkernel option is specified, but unfortunately EFI boot services may have already reserved some sub-1MB memory regions before the crash_reserve_low_1M() is called, which makes that the crash_reserve_low_1M() only own the remaining sub-1MB memory regions, not all sub-1MB memory regions, because, subsequently EFI boot services will free its own sub-1MB memory regions. Eventually, DMA will be able to allocate memory from the sub-1MB area and cause the following error:

Quoted 5 lines

So this patch is fixing a problem found in crash utility. We ever met the similar issue, later fixed by always reserving low 1M in commit 6f599d84231fd27 ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified"). Seems the commit is not fixing it completely.

Maybe I should add the "Fixes: 6f599d84231f" in front of 'Signed-off-by' as below:

Fixes 6f599d84231f ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified")
Quoted 24 lines
Quoted 9 lines

crash> kmem -s |grep invalid
kmem: dma-kmalloc-512: slab: ffffd52c40001900 invalid freepointer: ffff9403c0067300
kmem: dma-kmalloc-512: slab: ffffd52c40001900 invalid freepointer: ffff9403c0067300 crash> vtop ffff9403c0067300
VIRTUAL PHYSICAL
ffff9403c0067300 67300 --->The physical address falls into this range [0x0000000000063000-0x000000000008efff]

kernel debugging log:
...

Quoted 13 lines
[    0.008927] memblock_reserve: [0x0000000000010000-0x0000000000013fff] efi_reserve_boot_services+0x85/0xd0
[    0.008930] memblock_reserve: [0x0000000000063000-0x000000000008efff] efi_reserve_boot_services+0x85/0xd0
...
[    0.009425] memblock_reserve: [0x0000000000000000-0x00000000000fffff] crash_reserve_low_1M+0x2c/0x49
...
[    0.010586] Zone ranges:
[    0.010587]   DMA      [mem 0x0000000000001000-0x0000000000ffffff]
[    0.010589]   DMA32    [mem 0x0000000001000000-0x00000000ffffffff]
[    0.010591]   Normal   [mem 0x0000000100000000-0x0000000c7fffffff]
[    0.010593]   Device   empty
...
[    8.814894] __memblock_free_late: [0x0000000000063000-0x000000000008efff] efi_free_boot_services+0x14b/0x23b
[    8.815793] __memblock_free_late: [0x0000000000010000-0x0000000000013fff] efi_free_boot_services+0x14b/0x23b
Quoted 8 lines

In commit 6f599d84231fd27, we call crash_reserve_low_1M() to lock the whole low 1M area if crashkernel is specified in kernel cmdline. But earlier efi_reserve_boot_services() invokation will break the intention of the whole low 1M reserving. In efi_reserve_boot_services(), if any memory under low 1M hasn't been reserved, it will call memblock_reserve() to reserve it and leave it to efi_free_boot_services() to free.

Good understanding.

Quoted 4 lines

Hi Lianbo,

Please correct me if I am wrong or anything is missed. IIUC, can we move efi_reserve_boot_services() after reserve_real_mode() to fix this bug?

What do you think about the following changes?

patch [1]:

arch/x86/kernel/setup.c +6 -6
index 5ecd69a48393..c343de3178ec 100644
@@ -1064,12 +1064,6 @@ void __init setup_arch(char **cmdline_p)
        efi_esrt_init();
        efi_mokvar_table_init();
 
-       /*
-        * The EFI specification says that boot service code won't be
-        * called after ExitBootServices(). This is, in fact, a lie.
-        */
-       efi_reserve_boot_services();
-
        /* preallocate 4k for mptable mpc */
        e820__memblock_alloc_reserved_mpc_new();
 
@@ -1087,6 +1081,12 @@ void __init setup_arch(char **cmdline_p)
        trim_platform_memory_ranges();
        trim_low_memory_range();
 
+       /*
+        * The EFI specification says that boot service code won't be
+        * called after ExitBootServices(). This is, in fact, a lie.
+        */
+       efi_reserve_boot_services();
+
        init_mem_mapping();
 
        idt_setup_early_pf();
Quoted 2 lines

Or move reserve_real_mode() before efi_reserve_boot_services() since those real mode regions are all under 1M? Assume efi boot code/data

Or patch [2]

arch/x86/kernel/setup.c +1 -2
index 5ecd69a48393..ceec5af0dfab 100644
@@ -1058,6 +1058,7 @@ void __init setup_arch(char **cmdline_p)
        sev_setup_arch();
 
        reserve_bios_regions();
+       reserve_real_mode();
 
        efi_fake_memmap();
        efi_find_mirror();
@@ -1082,8 +1083,6 @@ void __init setup_arch(char **cmdline_p)
                        (max_pfn_mapped<<PAGE_SHIFT) - 1);
 #endif
 
-       reserve_real_mode();
-
        trim_platform_memory_ranges();
        trim_low_memory_range();
 

I tested the above two changes on my machine, both work well.

But I'm not very sure if the reordering of the code may affect the startup(boot) on other machines. Any comments about this?

Thanks.
Lianbo

Quoted 50 lines

won't rely on low 1M area any more at this moment.

Thanks
Baoquan

Quoted 45 lines

Do not release sub-1MB memory regions even though they are reserved by EFI boot services, so that always reserve all sub-1MB memory regions when the crashkernel option is specified.

---
 arch/x86/platform/efi/quirks.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
arch/x86/platform/efi/quirks.c +14 -0
index 67d93a243c35..637f932c4fd4 100644
@@ -18,6 +18,7 @@
 #include <asm/cpu_device_id.h>
 #include <asm/realmode.h>
 #include <asm/reboot.h>
+#include <asm/cmdline.h>
 
 #define EFI_MIN_RESERVE 5120
 
@@ -303,6 +304,19 @@ void __init efi_arch_mem_reserve(phys_addr_t addr, u64 size)
  */
 static __init bool can_free_region(u64 start, u64 size)
 {
+	/*
+	 * Some sub-1MB memory regions may be reserved by EFI boot
+	 * services, and these memory regions will be released later
+	 * in the efi_free_boot_services().
+	 *
+	 * Do not release sub-1MB memory regions even though they are
+	 * reserved by EFI boot services, because, always reserve all
+	 * sub-1MB memory when the crashkernel option is specified.
+	 */
+	if (cmdline_find_option(boot_command_line, "crashkernel", NULL, 0) > 0
+		&& (start + size < (1<<20)))
+		return false;
+
 	if (start + size > __pa_symbol(_text) && start <= __pa_symbol(_end))
 		return false;
 
Signature · 1 line
2.17.1

A message that is not in the archive (sent off-list or not yet received): [email protected]

#4 lore raw
Quoted 1 line · H. Peter Anvin

Why don't we do this unconditionally? At the very best we gain half a megabyte of memory (except the trampoline, which has to live there, but it is only a few kilobytes.)

This is a great suggestion, thanks. I think we can fix it in this way to make code simpler. Then the specific caring of real mode in efi_free_boot_services() can be removed too.

Thanks
Baoquan

#5 lore raw
Quoted 11 lines · Baoquan He
Quoted 1 line · H. Peter Anvin

Why don't we do this unconditionally? At the very best we gain half a megabyte of memory (except the trampoline, which has to live there, but it is only a few kilobytes.)

This is a great suggestion, thanks. I think we can fix it in this way to make code simpler. Then the specific caring of real mode in efi_free_boot_services() can be removed too.

This whole situation makes me think that the code is buggy before and buggy after.

The issue here (I think) is that various pieces of code want to reserve specific pieces of otherwise-available low memory for their own nefarious uses. I don’t know *why* crash kernel needs this, but that doesn’t matter too much.

I propose that the right solution is to give low-memory-reserving code paths two chances to do what they need: once at the very beginning and once after EFI boot services are freed.

Alternatively, just reserve *all* otherwise unused sub 1M memory up front, then release it right after releasing boot services, and then invoke the special cases exactly once.

In either case, the result is that the crashkernel mess gets unified with the trampoline mess. One way the result is called twice and needs to be more careful, and the other way it’s called only once.

Just skipping freeing boot services seems wrong. It doesn’t unmap boot services, and skipping that is incorrect, I think. And it seems to result in a bogus memory map in which the system thinks that some crashkernel memory is EFI memory instead.

Let’s please just fix the problem instead of papering over it with more hacks.

Quoted 3 lines

Thanks
Baoquan

#6 lore raw
Quoted 7 lines · Andy Lutomirski
Quoted 7 lines · Baoquan He
Quoted 1 line · H. Peter Anvin

Why don't we do this unconditionally? At the very best we gain half a megabyte of memory (except the trampoline, which has to live there, but it is only a few kilobytes.)

Quoted 8 lines
Quoted 4 lines

This is a great suggestion, thanks. I think we can fix it in this way to make code simpler. Then the specific caring of real mode in efi_free_boot_services() can be removed too.

This whole situation makes me think that the code is buggy before and buggy after.

The issue here (I think) is that various pieces of code want to reserve specific pieces of otherwise-available low memory for their own nefarious uses. I don’t know *why* crash kernel needs this, but that doesn’t matter too much.

Kdump kernel also need go through real mode code path during bootup. It is not different than normal kernel except that it skips the firmware resetting. So kdump kernel needs low 1M as system RAM just as normal kernel does. Here we reserve the whole low 1M with memblock_reserve() to avoid any later kernel or driver data reside in this area. Otherwise, we need dump the content of this area to vmcore. As we know, when crash happened, the old memory of 1st kernel should be untouched until vmcore dumping read out its content. Meanwhile, kdump kernel need reuse low 1M. In the past, we used a back up region to copy out the low 1M area, and map the back up region into the low 1M area in vmcore elf file. In 6f599d84231fd27 ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified"), we changed to lock the whole low 1M to avoid writting any kernel data into, like this we can skip this area when dumping vmcore.

Above is why we try to memblock reserve the whole low 1M. We don't want to use it, just don't want anyone to use it in 1st kernel.

Quoted 4 lines

I propose that the right solution is to give low-memory-reserving code paths two chances to do what they need: once at the very beginning and once after EFI boot services are freed.

Alternatively, just reserve *all* otherwise unused sub 1M memory up front, then release it right after releasing boot services, and then invoke the special cases exactly once.

I am not sure if I got both suggested ways clearly. They look a little complicated in our case. As I explained at above, we want the whole low 1M locked up, not one piece or some pieces of it.

Quoted 4 lines

In either case, the result is that the crashkernel mess gets unified with the trampoline mess. One way the result is called twice and needs to be more careful, and the other way it’s called only once.

Just skipping freeing boot services seems wrong. It doesn’t unmap boot services, and skipping that is incorrect, I think. And it seems to result in a bogus memory map in which the system thinks that some crashkernel memory is EFI memory instead.

I like hpa's thought to lock the whole low 1M unconditionally since only a few KB except of trampoline area is there. Rethinking about it, doing it in can_free_region() may be risky because efi memory region could cross the 1M boundary, e.g [640K, 100M] with type of EFI_BOOT_SERVICES_CODE|EFI_BOOT_SERVICES_DATA, it could cause loss of memory. Just a wild guess, not very sure if the 1M boundary corssing can really happen. efi_reserve_boot_services() won't split regions.

If moving efi_reserve_boot_services() after reserve_real_mode() is not accepted, maybe we can call efi_mem_reserve(0, 1M) just as efi_esrt_init() has done.

_______________________________________________
kexec mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kexec

author
#7 lore raw

Thank you for the comment, H. Peter, Andy and Baoquan.

Quoted 32 lines · Baoquan He
Quoted 7 lines · Andy Lutomirski
Quoted 7 lines · Baoquan He
Quoted 1 line · H. Peter Anvin

Why don't we do this unconditionally? At the very best we gain half a megabyte of memory (except the trampoline, which has to live there, but it is only a few kilobytes.)

Quoted 8 lines
Quoted 4 lines

This is a great suggestion, thanks. I think we can fix it in this way to make code simpler. Then the specific caring of real mode in efi_free_boot_services() can be removed too.

This whole situation makes me think that the code is buggy before and buggy after.

The issue here (I think) is that various pieces of code want to reserve specific pieces of otherwise-available low memory for their own nefarious uses. I don’t know *why* crash kernel needs this, but that doesn’t matter too much.

Kdump kernel also need go through real mode code path during bootup. It is not different than normal kernel except that it skips the firmware resetting. So kdump kernel needs low 1M as system RAM just as normal kernel does. Here we reserve the whole low 1M with memblock_reserve() to avoid any later kernel or driver data reside in this area. Otherwise, we need dump the content of this area to vmcore. As we know, when crash happened, the old memory of 1st kernel should be untouched until vmcore dumping read out its content. Meanwhile, kdump kernel need reuse low 1M. In the past, we used a back up region to copy out the low 1M area, and map the back up region into the low 1M area in vmcore elf file. In 6f599d84231fd27 ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified"), we changed to lock the whole low 1M to avoid writting any kernel data into, like this we can skip this area when dumping vmcore.

Quoted 8 lines

Above is why we try to memblock reserve the whole low 1M. We don't want to use it, just don't want anyone to use it in 1st kernel.

Quoted 4 lines

I propose that the right solution is to give low-memory-reserving code paths two chances to do what they need: once at the very beginning and once after EFI boot services are freed.

Alternatively, just reserve *all* otherwise unused sub 1M memory up front, then release it right after releasing boot services, and then invoke the special cases exactly once.

After EFI boot services are freed, I'm worried that it's a bit late. All sub-1M memory regions need to be reserved early as soon as possible.

Quoted 7 lines

I am not sure if I got both suggested ways clearly. They look a little complicated in our case. As I explained at above, we want the whole low 1M locked up, not one piece or some pieces of it.

Quoted 3 lines

In either case, the result is that the crashkernel mess gets unified with the trampoline mess. One way the result is called twice and needs to be more careful, and the other way it’s called only once.

That may still have a chance to allocate memory from sub-1M regions at some point, because EFI boot services will be freed after EFI enters virtual mode, it looks late.

Quoted 7 lines
Quoted 1 line

Just skipping freeing boot services seems wrong. It doesn’t unmap boot services, and skipping that is incorrect, I think. And it seems to result in a bogus memory map in which the system thinks that some crashkernel memory is EFI memory instead.

I like hpa's thought to lock the whole low 1M unconditionally since only a few KB except of trampoline area is there. Rethinking about it, doing it in can_free_region() may be risky because efi memory region could cross the 1M boundary, e.g [640K, 100M] with type of EFI_BOOT_SERVICES_CODE|EFI_BOOT_SERVICES_DATA, it could cause loss of memory.

Theoretically, yes. But so far I haven't seen the situation of crossing the 1M boundary.

Thanks.
Lianbo

Quoted 7 lines

Just a wild guess, not very sure if the 1M boundary corssing can really happen. efi_reserve_boot_services() won't split regions.

If moving efi_reserve_boot_services() after reserve_real_mode() is not accepted, maybe we can call efi_mem_reserve(0, 1M) just as efi_esrt_init() has done.

#8 lore raw
In reply to Baoquan He
Quoted 37 lines · Baoquan He
Quoted 7 lines · Andy Lutomirski
Quoted 7 lines · Baoquan He
Quoted 1 line · H. Peter Anvin

Why don't we do this unconditionally? At the very best we gain half a megabyte of memory (except the trampoline, which has to live there, but it is only a few kilobytes.)

Quoted 8 lines
Quoted 4 lines

This is a great suggestion, thanks. I think we can fix it in this way to make code simpler. Then the specific caring of real mode in efi_free_boot_services() can be removed too.

This whole situation makes me think that the code is buggy before and buggy after.

The issue here (I think) is that various pieces of code want to reserve specific pieces of otherwise-available low memory for their own nefarious uses. I don’t know *why* crash kernel needs this, but that doesn’t matter too much.

Kdump kernel also need go through real mode code path during bootup. It is not different than normal kernel except that it skips the firmware resetting. So kdump kernel needs low 1M as system RAM just as normal kernel does. Here we reserve the whole low 1M with memblock_reserve() to avoid any later kernel or driver data reside in this area. Otherwise, we need dump the content of this area to vmcore. As we know, when crash happened, the old memory of 1st kernel should be untouched until vmcore dumping read out its content. Meanwhile, kdump kernel need reuse low 1M. In the past, we used a back up region to copy out the low 1M area, and map the back up region into the low 1M area in vmcore elf file. In 6f599d84231fd27 ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified"), we changed to lock the whole low 1M to avoid writting any kernel data into, like this we can skip this area when dumping vmcore.

Above is why we try to memblock reserve the whole low 1M. We don't want to use it, just don't want anyone to use it in 1st kernel.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

I propose that the right solution is to give low-memory-reserving code paths two chances to do what they need: once at the very beginning and once after EFI boot services are freed.

Alternatively, just reserve *all* otherwise unused sub 1M memory up front, then release it right after releasing boot services, and then invoke the special cases exactly once.

I am not sure if I got both suggested ways clearly. They look a little complicated in our case. As I explained at above, we want the whole low 1M locked up, not one piece or some pieces of it.

My second suggestion is probably the better one. Here it is, concretely:

The early (pre-free_efi_boot_services) code just reserves all available sub-1M memory unconditionally, but it specially marks it as reserved-but-available-later. We stop allocating the trampoline page at this stage.

In free_efi_boot_services, instead of *freeing* the sub-1M memory, we stick it in the pile of reserved memory created in the early step. This may involve splitting a block, kind of like the current trampoline late allocation works.

Then, *after* free_efi_boot_services(), we run a single block of code that lets everything that wants sub-1M code claim some. This means that the trampoline gets allocated and, if crashkernel wants to claim everything else, it can. After that, everything still unclaimed gets freed.

Does that make sense?

--Andy

_______________________________________________
kexec mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kexec

#9 lore raw
Quoted 58 lines · Andy Lutomirski
Quoted 37 lines · Baoquan He
Quoted 7 lines · Andy Lutomirski
Quoted 7 lines · Baoquan He
Quoted 1 line · H. Peter Anvin

Why don't we do this unconditionally? At the very best we gain half a megabyte of memory (except the trampoline, which has to live there, but it is only a few kilobytes.)

Quoted 8 lines
Quoted 4 lines

This is a great suggestion, thanks. I think we can fix it in this way to make code simpler. Then the specific caring of real mode in efi_free_boot_services() can be removed too.

This whole situation makes me think that the code is buggy before and buggy after.

The issue here (I think) is that various pieces of code want to reserve specific pieces of otherwise-available low memory for their own nefarious uses. I don’t know *why* crash kernel needs this, but that doesn’t matter too much.

Kdump kernel also need go through real mode code path during bootup. It is not different than normal kernel except that it skips the firmware resetting. So kdump kernel needs low 1M as system RAM just as normal kernel does. Here we reserve the whole low 1M with memblock_reserve() to avoid any later kernel or driver data reside in this area. Otherwise, we need dump the content of this area to vmcore. As we know, when crash happened, the old memory of 1st kernel should be untouched until vmcore dumping read out its content. Meanwhile, kdump kernel need reuse low 1M. In the past, we used a back up region to copy out the low 1M area, and map the back up region into the low 1M area in vmcore elf file. In 6f599d84231fd27 ("x86/kdump: Always reserve the low 1M when the crashkernel option is specified"), we changed to lock the whole low 1M to avoid writting any kernel data into, like this we can skip this area when dumping vmcore.

Above is why we try to memblock reserve the whole low 1M. We don't want to use it, just don't want anyone to use it in 1st kernel.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

I propose that the right solution is to give low-memory-reserving code paths two chances to do what they need: once at the very beginning and once after EFI boot services are freed.

Alternatively, just reserve *all* otherwise unused sub 1M memory up front, then release it right after releasing boot services, and then invoke the special cases exactly once.

I am not sure if I got both suggested ways clearly. They look a little complicated in our case. As I explained at above, we want the whole low 1M locked up, not one piece or some pieces of it.

My second suggestion is probably the better one. Here it is, concretely:

The early (pre-free_efi_boot_services) code just reserves all available sub-1M memory unconditionally, but it specially marks it as reserved-but-available-later. We stop allocating the trampoline page at this stage.

In free_efi_boot_services, instead of *freeing* the sub-1M memory, we stick it in the pile of reserved memory created in the early step. This may involve splitting a block, kind of like the current trampoline late allocation works.

Quoted 5 lines

Then, *after* free_efi_boot_services(), we run a single block of code that lets everything that wants sub-1M code claim some. This means that the trampoline gets allocated and, if crashkernel wants to claim everything else, it can. After that, everything still unclaimed gets freed.

void __init setup_arch(char **cmdline_p)
{
...

	efi_reserve_boot_services();
	e820__memblock_alloc_reserved_mpc_new();
#ifdef CONFIG_X86_CHECK_BIOS_CORRUPTION
        setup_bios_corruption_check();
#endif
        reserve_real_mode();                                                                                                                      

        trim_platform_memory_ranges();
        trim_low_memory_range();

...
}

After efi_reserve_boot_services(), there are several function calling to require memory reservation under low 1M.

asmlinkage __visible void __init __no_sanitize_address start_kernel(void) {
...

	setup_arch(&command_line);
...
	mm_init();
		--> mem_init();
			 -->memblock_free_all();

...

#ifdef CONFIG_X86
        if (efi_enabled(EFI_RUNTIME_SERVICES))
                efi_enter_virtual_mode();
			-->efi_free_boot_services();
				-->memblock_free_late();

#endif
...
}

So from the code flow, we can see that buddy allocator is built in mm_init() which puts all memory from memblock.memory excluding memblock.reserved into buddy. And much later, we call efi_free_boot_services() to release those reserved efi boot memory into buddy too.

Are you suggesting we should do the memory reservation from low 1M after efi_free_boot_services()? To require memory pages from buddy for them? Please help point out my misunderstanding if have any.

With my understanding, in non-efi case, we have done the memory reservation with memblock_reserve(), e.g e820__memblock_alloc_reserved_mpc_new, reserve_real_mode() are calling to do. Just efi_reserve|free_boot_services() break them when efi is enabled. We can do them again in efi_free_boot_services() just like the real_mode reservation does.

Thanks
Baoquan

_______________________________________________
kexec mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kexec

A message that is not in the archive (sent off-list or not yet received): CANU+ZydgWTSg+iUix8ggn-cSPpg8qtShaUQ47cOeeMxFmXp_zQ@mail.gmail.com

A message that is not in the archive (sent off-list or not yet received): CANU+ZydyKsctuFjPfBz7PuS=FaUtK0gs5Lq06pL5nuRJKe+J0w@mail.gmail.com

#12 lore raw

Hi Lianbo,

Quoted 166 lines · lijiang

Also add mail lists and more people in the cc list.

Thanks.
Lianbo

Quoted 158 lines · lijiang

Hi, Baoquan, Andy
Sorry for the late reply.

Quoted 132 lines · Baoquan He
Quoted 9 lines · Andy Lutomirski
Quoted 8 lines · Baoquan He
Quoted 6 lines · Andy Lutomirski
Quoted 6 lines · Baoquan He
Quoted 1 line · H. Peter Anvin

Why don't we do this unconditionally? At the very best we gain

half a megabyte of memory (except the trampoline, which has to live there, but it is only a few kilobytes.)

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

This is a great suggestion, thanks. I think we can fix it in this

way to

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 3 lines

make code simpler. Then the specific caring of real mode in efi_free_boot_services() can be removed too.

This whole situation makes me think that the code is buggy before

and buggy after.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

The issue here (I think) is that various pieces of code want to

reserve specific pieces of otherwise-available low memory for their own nefarious uses. I don’t know *why* crash kernel needs this, but that doesn’t matter too much.

Quoted 2 lines
Quoted 2 lines

Kdump kernel also need go through real mode code path during bootup.

It

Quoted 4 lines
Quoted 4 lines

is not different than normal kernel except that it skips the firmware resetting. So kdump kernel needs low 1M as system RAM just as normal kernel does. Here we reserve the whole low 1M with memblock_reserve() to avoid any later kernel or driver data reside in this area.

Otherwise,

Quoted 1 line
Quoted 1 line

we need dump the content of this area to vmcore. As we know, when

crash

Quoted 1 line
Quoted 1 line

happened, the old memory of 1st kernel should be untouched until

vmcore

Quoted 1 line
Quoted 1 line

dumping read out its content. Meanwhile, kdump kernel need reuse low

1M.

Quoted 3 lines
Quoted 3 lines

In the past, we used a back up region to copy out the low 1M area, and map the back up region into the low 1M area in vmcore elf file. In 6f599d84231fd27 ("x86/kdump: Always reserve the low 1M when the

crashkernel

Quoted 5 lines
Quoted 5 lines

option is specified"), we changed to lock the whole low 1M to avoid writting any kernel data into, like this we can skip this area when dumping vmcore.

Above is why we try to memblock reserve the whole low 1M. We don't

want

Quoted 4 lines
Quoted 4 lines

to use it, just don't want anyone to use it in 1st kernel.

Quoted 2 lines

I propose that the right solution is to give low-memory-reserving

code paths two chances to do what they need: once at the very beginning and once after EFI boot services are freed.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

Alternatively, just reserve *all* otherwise unused sub 1M memory up

front, then release it right after releasing boot services, and then invoke the special cases exactly once.

Quoted 3 lines
Quoted 3 lines

I am not sure if I got both suggested ways clearly. They look a little complicated in our case. As I explained at above, we want the whole

low

Quoted 3 lines
Quoted 1 line

1M locked up, not one piece or some pieces of it.

My second suggestion is probably the better one. Here it is,

concretely:

Quoted 11 lines

The early (pre-free_efi_boot_services) code just reserves all available sub-1M memory unconditionally, but it specially marks it as reserved-but-available-later. We stop allocating the trampoline page at this stage.

In free_efi_boot_services, instead of *freeing* the sub-1M memory, we stick it in the pile of reserved memory created in the early step. This may involve splitting a block, kind of like the current trampoline late allocation works.

Quoted 5 lines

Then, *after* free_efi_boot_services(), we run a single block of code that lets everything that wants sub-1M code claim some. This means that the trampoline gets allocated and, if crashkernel wants to claim everything else, it can. After that, everything still unclaimed gets freed.

void __init setup_arch(char **cmdline_p)
{
...

        efi_reserve_boot_services();
        e820__memblock_alloc_reserved_mpc_new();
#ifdef CONFIG_X86_CHECK_BIOS_CORRUPTION
        setup_bios_corruption_check();
#endif
        reserve_real_mode();


        trim_platform_memory_ranges();
        trim_low_memory_range();

...
}

After efi_reserve_boot_services(), there are several function calling to require memory reservation under low 1M.

asmlinkage __visible void __init __no_sanitize_address start_kernel(void)

{
...

        setup_arch(&command_line);
...
        mm_init();
                --> mem_init();
                         -->memblock_free_all();

...

#ifdef CONFIG_X86
        if (efi_enabled(EFI_RUNTIME_SERVICES))
                efi_enter_virtual_mode();
                        -->efi_free_boot_services();
                                -->memblock_free_late();

#endif
...
}

So from the code flow, we can see that buddy allocator is built in mm_init() which puts all memory from memblock.memory excluding memblock.reserved into buddy. And much later, we call efi_free_boot_services() to release those reserved efi boot memory into buddy too.

Quoted 9 lines

Are you suggesting we should do the memory reservation from low 1M after efi_free_boot_services()? To require memory pages from buddy for them? Please help point out my misunderstanding if have any.

With my understanding, in non-efi case, we have done the memory reservation with memblock_reserve(), e.g e820__memblock_alloc_reserved_mpc_new, reserve_real_mode() are calling to do. Just efi_reserve|free_boot_services() break them when efi is

Yes. But Andy also suggested to reserve all available sub-1M memory
**unconditionally**
in the early code.

enabled. We can do them again in efi_free_boot_services() just like the

Quoted 2 lines

real_mode reservation does.

Do you mean to call the memblock_reserve() in
the efi_free_boot_services()? Or anything else?
Would you mind sharing more details about this?

Hmm, maybe no. There's no chance to call memblock_reserve() after memblock_free_late().

I suggested making change in efi_free_boot_services() because the similar problem has been encountered by Andy himself and fixed in efi_free_boot_services(). Please check below commit for reference. The patch described the phenemenon, while explained why in code comment.

commit 5bc653b7318217c54244a14f248f1f07abe0a865

Author: Andy Lutomirski <[email protected]>
Date:   Wed Aug 10 02:29:17 2016 -0700

x86/efi: Allocate a trampoline if needed in efi_free_boot_services()

So we could handle it in the same place by extending the area to the whole low-1M unconditionally. This is 1st way of three I can think of.

The other two are:
2) Move efi_reserve_boot_services() down to be above init_mem_mapping(). until init_mem_mapping(), we stop reserving memory from low-1M with memblock_reserve(). The change is like below:

arch/x86/kernel/setup.c +2 -1
index 72920af0b3c0..93b2106d2050 100644
@@ -1067,7 +1067,6 @@ void __init setup_arch(char **cmdline_p)
 	 * The EFI specification says that boot service code won't be
 	 * called after ExitBootServices(). This is, in fact, a lie.
 	 */
-	efi_reserve_boot_services();
 
 	/* preallocate 4k for mptable mpc */
 	e820__memblock_alloc_reserved_mpc_new();
@@ -1090,6 +1089,8 @@ void __init setup_arch(char **cmdline_p)
 	 */
 	trim_snb_memory();
 
+	efi_reserve_boot_services();
+
 	init_mem_mapping();
 
 	idt_setup_early_pf();

3) Do it in efi_reserve_boot_services() as you are doing in this patch.

I prefer the 1st and 2nd way.

And by the way, the original code of commit 5bc653b731821("x86/efi: Allocate a trampoline if needed in efi_free_boot_services()") could be buggy. I remember you ever pasted the boot log of system where you reproduced issue and tested the tested patch, there are three separate pages from low-1M reserved by boot services. If any of them falls into real mode area, it will break the fix of commit 5bc653b731821.

Thanks
Baoquan

Quoted 23 lines
Quoted 23 lines
arch/x86/platform/efi/quirks.c +2 -0
index 7850111008a8..d02f12a60457 100644
@@ -453,6 +453,8 @@ void __init efi_free_boot_services(void)
                memblock_free_late(start, size);
        }
 
+       memblock_reserve(0, 1<<20);
+
        if (!num_entries)
                return;
 

Thanks.
Lianbo

Thanks

Quoted 2 lines

Baoquan