[PATCH] Revert "Revert "kdump,

Discussion:

Thomas Garnier

2016-12-15 16:17:28 UTC

This reverts commit 49fd897573c97b0eaf10f47d850027d78c456cd7.

Reverting back to commit 0549a3c because the values are used by crash
and other tools already. I expected this commit would not go through given
the unresolved comments. I want it to be easy to resolve major memory
section positions when KASLR memory randomization is enabled.

Signed-off-by: Thomas Garnier <***@google.com>
---
arch/x86/kernel/machine_kexec_64.c | 3 +++
include/linux/kexec.h | 6 ++++++
2 files changed, 9 insertions(+)

diff --git a/arch/x86/kernel/machine_kexec_64.c b/arch/x86/kernel/machine_kexec_64.c
index 2e3c34b..05f3367 100644
--- a/arch/x86/kernel/machine_kexec_64.c
+++ b/arch/x86/kernel/machine_kexec_64.c
@@ -339,6 +339,9 @@ void arch_crash_save_vmcoreinfo(void)
kaslr_offset());
VMCOREINFO_NUMBER(KERNEL_IMAGE_SIZE);
VMCOREINFO_PHYS_BASE(phys_base);
+ VMCOREINFO_PAGE_OFFSET(PAGE_OFFSET);
+ VMCOREINFO_VMALLOC_START(VMALLOC_START);
+ VMCOREINFO_VMEMMAP_START(VMEMMAP_START);
}

/* arch-dependent functionality related to kexec file-based syscall */
diff --git a/include/linux/kexec.h b/include/linux/kexec.h
index e98e546..ff9c876 100644
--- a/include/linux/kexec.h
+++ b/include/linux/kexec.h
@@ -285,6 +285,12 @@ phys_addr_t paddr_vmcoreinfo_note(void);
vmcoreinfo_append_str("CONFIG_%s=y\n", #name)
#define VMCOREINFO_PHYS_BASE(value) \
vmcoreinfo_append_str("PHYS_BASE=%lx\n", (unsigned long)value)
+#define VMCOREINFO_PAGE_OFFSET(value) \
+ vmcoreinfo_append_str("PAGE_OFFSET=%lx\n", (unsigned long)value)
+#define VMCOREINFO_VMALLOC_START(value) \
+ vmcoreinfo_append_str("VMALLOC_START=%lx\n", (unsigned long)value)
+#define VMCOREINFO_VMEMMAP_START(value) \
+ vmcoreinfo_append_str("VMEMMAP_START=%lx\n", (unsigned long)value)

extern struct kimage *kexec_image;
extern struct kimage *kexec_crash_image;

--
2.8.0.rc3.226.g39d4020

Eric W. Biederman

2016-12-15 17:50:31 UTC

Permalink

Post by Thomas Garnier
This reverts commit 49fd897573c97b0eaf10f47d850027d78c456cd7.
Reverting back to commit 0549a3c because the values are used by crash
and other tools already. I expected this commit would not go through given
the unresolved comments. I want it to be easy to resolve major memory
section positions when KASLR memory randomization is enabled.

This patch is broken. The commit referenced is wrong,
as is the justification.

These values are not in fact widely used by userspace (they are brand new).

This is a very fragile approach relying on kernel implementation
details, so if we can do anything else that is more robust it
is much more likely to pass the test of time.

And yes a more robust implementation has been already discussed.

Post by Thomas Garnier
---
arch/x86/kernel/machine_kexec_64.c | 3 +++
include/linux/kexec.h | 6 ++++++
2 files changed, 9 insertions(+)
diff --git a/arch/x86/kernel/machine_kexec_64.c b/arch/x86/kernel/machine_kexec_64.c
index 2e3c34b..05f3367 100644
--- a/arch/x86/kernel/machine_kexec_64.c
+++ b/arch/x86/kernel/machine_kexec_64.c
@@ -339,6 +339,9 @@ void arch_crash_save_vmcoreinfo(void)
kaslr_offset());
VMCOREINFO_NUMBER(KERNEL_IMAGE_SIZE);
VMCOREINFO_PHYS_BASE(phys_base);
+ VMCOREINFO_PAGE_OFFSET(PAGE_OFFSET);
+ VMCOREINFO_VMALLOC_START(VMALLOC_START);
+ VMCOREINFO_VMEMMAP_START(VMEMMAP_START);
}
/* arch-dependent functionality related to kexec file-based syscall */
diff --git a/include/linux/kexec.h b/include/linux/kexec.h
index e98e546..ff9c876 100644
--- a/include/linux/kexec.h
+++ b/include/linux/kexec.h
@@ -285,6 +285,12 @@ phys_addr_t paddr_vmcoreinfo_note(void);
vmcoreinfo_append_str("CONFIG_%s=y\n", #name)
#define VMCOREINFO_PHYS_BASE(value) \
vmcoreinfo_append_str("PHYS_BASE=%lx\n", (unsigned long)value)
+#define VMCOREINFO_PAGE_OFFSET(value) \
+ vmcoreinfo_append_str("PAGE_OFFSET=%lx\n", (unsigned long)value)
+#define VMCOREINFO_VMALLOC_START(value) \
+ vmcoreinfo_append_str("VMALLOC_START=%lx\n", (unsigned long)value)
+#define VMCOREINFO_VMEMMAP_START(value) \
+ vmcoreinfo_append_str("VMEMMAP_START=%lx\n", (unsigned long)value)
extern struct kimage *kexec_image;
extern struct kimage *kexec_crash_image;

Thomas Garnier

2016-12-15 18:16:38 UTC

Permalink

On Thu, Dec 15, 2016 at 9:50 AM, Eric W. Biederman

Post by Eric W. Biederman

This patch is broken. The commit referenced is wrong,

Yes, I based them on linux-next. I can update them to linux main tree.
Sorry about that.

Post by Eric W. Biederman
as is the justification.
These values are not in fact widely used by userspace (they are brand new).

They were new and got through to the master tree before and on 4.9. I
didn't get any feedback on that when it went through.

Post by Eric W. Biederman
This is a very fragile approach relying on kernel implementation
details, so if we can do anything else that is more robust it
is much more likely to pass the test of time.
And yes a more robust implementation has been already discussed.

There were discussed for the PAGE_OFFSET but not VMALLOC_START. What's
your approach for it?

Post by Eric W. Biederman

--
Thomas

Thomas Garnier

2016-12-15 19:45:10 UTC

Permalink

Post by Thomas Garnier
On Thu, Dec 15, 2016 at 9:50 AM, Eric W. Biederman

Post by Eric W. Biederman

This patch is broken. The commit referenced is wrong,

Yes, I based them on linux-next. I can update them to linux main tree.
Sorry about that.

Post by Eric W. Biederman
as is the justification.
These values are not in fact widely used by userspace (they are brand new).

They were new and got through to the master tree before and on 4.9. I
didn't get any feedback on that when it went through.

There were discussed for the PAGE_OFFSET but not VMALLOC_START. What's
your approach for it?

Also with improvement on KASLR memory randomization, I don't think we
should assume memory sections by looking at PT_LOAD first entries. We
will certainly try to randomize the order. That's why I thought
clearly indicating which VA define which memory section was a good
idea.

I also want to find a time proof approach where we can use it no
matter the changes to KASLR.

Post by Thomas Garnier

Post by Eric W. Biederman

--
Thomas

Baoquan He

2016-12-16 07:52:24 UTC

Permalink

Post by Thomas Garnier

Post by Eric W. Biederman
as is the justification.
These values are not in fact widely used by userspace (they are brand new).

They were new and got through to the master tree before and on 4.9. I
didn't get any feedback on that when it went through.

There were discussed for the PAGE_OFFSET but not VMALLOC_START. What's
your approach for it?

Sorry, Thomas. I didn't realize these are also required other than
Makedumpfile. So just revert it after we use a new way to get
page_offset and translate VA to PA by page table in makedumpfile. For
the time being they are not needed any more by makedumpfile.

While from below discussion we can see Atsushi, maintainer of
Makedumpfile, still worry about the parsing PT_LOAD segment header way.
I personally think exporting them is easier and more understandable for
user space utility.

http://lists.infradead.org/pipermail/kexec/2016-December/017815.html

I can't think of other way for your tool, maybe page_offset you can
use the similar way makedumpfile is taking. As for VMALLOC_START and
VMEMMAP_START, it could be necessary. And if you are planning to
randomize the memory section order, thinking about them carefully
now is a good chance lest they need be adjusted again later.

Sorry again for this.

Post by Thomas Garnier
Also with improvement on KASLR memory randomization, I don't think we
should assume memory sections by looking at PT_LOAD first entries. We
will certainly try to randomize the order. That's why I thought
clearly indicating which VA define which memory section was a good
idea.
I also want to find a time proof approach where we can use it no
matter the changes to KASLR.

Post by Thomas Garnier

Post by Eric W. Biederman

--
Thomas

--
Thomas
_______________________________________________
kexec mailing list
http://lists.infradead.org/mailman/listinfo/kexec