Skip to contentSkip to conversation

Kdump(x86): add note type NT_KDUMPINFO to kernel core dumps

PATCH In review Vivek Goyal 5 participants 21 messages Sep 21, 2005 – Sep 26, 2005 2 messages not in the archive

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Kdump(x86): add note type NT_KDUMPINFO to kernel core dumps

o This patch adds a new note type NT_KDUMPINFO. This note is added with kernel core dumps generated by kdump.

o This note mainly communicates the information required by kernel dump analysis tool "crash" to analyze the kernel core dump. As of now it contains the pointer to task struct of panicing task and page size. Page size is irrelevant for i386 but is required for architectures like ia64 and ppc64.

o gdb is not affected by this change as gdb need not to parse this note.

---

 linux-2.6.14-rc1-mm1-vivek/arch/i386/kernel/crash.c |    9 +++++++++
 linux-2.6.14-rc1-mm1-vivek/include/linux/elf.h      |    2 +-
 linux-2.6.14-rc1-mm1-vivek/include/linux/elfcore.h  |   13 +++++++++++++
 3 files changed, 23 insertions(+), 1 deletion(-)
arch/i386/kernel/crash.c renamed from arch/i386/kernel/crash.c~kdump-x86-add-note-type-nt-kdump +9 -0
diff -puN arch/i386/kernel/crash.c~kdump-x86-add-note-type-nt-kdump arch/i386/kernel/crash.c
@@ -62,6 +62,7 @@ static void final_note(u32 *buf)
 static void crash_save_this_cpu(struct pt_regs *regs, int cpu)
 {
 	struct elf_prstatus prstatus;
+	struct elf_kdumpinfo kdumpinfo;
 	u32 *buf;
 
 	if ((cpu < 0) || (cpu >= NR_CPUS))
@@ -80,6 +81,14 @@ static void crash_save_this_cpu(struct p
 	elf_core_copy_regs(&prstatus.pr_reg, regs);
 	buf = append_elf_note(buf, "CORE", NT_PRSTATUS, &prstatus,
 				sizeof(prstatus));
+
+	/* Also store NT_KDUMPINFO */
+	if (cpu == crashing_cpu) {
+		kdumpinfo.page_shift = PAGE_SHIFT;
+		kdumpinfo.panic_tsk = current;
+		buf = append_elf_note(buf, "CORE", NT_KDUMPINFO, &kdumpinfo,
+					sizeof(kdumpinfo));
+	}
 	final_note(buf);
 }
 
include/linux/elfcore.h renamed from include/linux/elfcore.h~kdump-x86-add-note-type-nt-kdump +13 -0
diff -puN include/linux/elfcore.h~kdump-x86-add-note-type-nt-kdump include/linux/elfcore.h
@@ -79,9 +79,22 @@ struct elf_prpsinfo
 	char	pr_psargs[ELF_PRARGSZ];	/* initial part of arg list */
 };
 
+/*
+ * NT_KDUMPINFO can be used to communicate additional information required for
+ * kernel core dumps. Additional information includes kernel configuration
+ * variables like page size and any other relevant data required by
+ * debugger (crash in this case).
+*/
+struct elf_kdumpinfo
+{
+	char page_shift;		/* Page size */
+	struct task_struct *panic_tsk;	/* Pointer to panic task_struct */
+};
+
 #ifndef __KERNEL__
 typedef struct elf_prstatus prstatus_t;
 typedef struct elf_prpsinfo prpsinfo_t;
+typedef struct elf_kdumpinfo kdumpinfo_t;
 #define PRARGSZ ELF_PRARGSZ 
 #endif
 
include/linux/elf.h renamed from include/linux/elf.h~kdump-x86-add-note-type-nt-kdump +1 -1
diff -puN include/linux/elf.h~kdump-x86-add-note-type-nt-kdump include/linux/elf.h
@@ -391,7 +391,7 @@ typedef struct elf64_shdr {
 #define NT_TASKSTRUCT	4
 #define NT_AUXV		6
 #define NT_PRXFPREG     0x46e62b7f      /* copied from gdb5.1/include/elf/common.h */
-
+#define NT_KDUMPINFO	7		/* Used for kernel core dumps */
 
 /* Note header in a PT_NOTE section */
 typedef struct elf32_note {

_

#1 lore raw
Quoted 9 lines · Vivek Goyal

o This patch adds a new note type NT_KDUMPINFO. This note is added with kernel core dumps generated by kdump.

o This note mainly communicates the information required by kernel dump analysis tool "crash" to analyze the kernel core dump. As of now it contains the pointer to task struct of panicing task and page size. Page size is irrelevant for i386 but is required for architectures like ia64 and ppc64.

o gdb is not affected by this change as gdb need not to parse this note.

A couple of things.
- The name of your note is terribly generic, so it seems a poor choice.

- Why do we need to capture the page size at the time of the crash? Isn't the page size a compile time option? Won't sys_getpagesize() get you this information before the crash? Why do we need the kernels page size at all?

- Why do you avoid storing the current task on the other cpus?

- Can't we derive the current task from the existing register information already captured. If not would a little extra debug information captured at compile time be better?

- You don't address the issue of architectural control registers. So you are going to need another note at some point. (Not necessarily a bad thing).

Quoted 19 lines
include/linux/elfcore.h renamed from include/linux/elfcore.h~kdump-x86-add-note-type-nt-kdump +13 -1
+/*
+ * NT_KDUMPINFO can be used to communicate additional information required for
+ * kernel core dumps. Additional information includes kernel configuration
+ * variables like page size and any other relevant data required by
+ * debugger (crash in this case).
+*/
+struct elf_kdumpinfo
+{
+	char page_shift;		/* Page size */
+	struct task_struct *panic_tsk;	/* Pointer to panic task_struct */
+};
+
 
 #define NT_TASKSTRUCT	4
 #define NT_AUXV		6
 #define NT_PRXFPREG 0x46e62b7f /* copied from gdb5.1/include/elf/common.h */
-
+#define NT_KDUMPINFO	7		/* Used for kernel core dumps */
#2 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Quoted 10 lines · Eric W. Biederman
Quoted 8 lines · Vivek Goyal

o This patch adds a new note type NT_KDUMPINFO. This note is added with kernel core dumps generated by kdump.

o This note mainly communicates the information required by kernel dump analysis tool "crash" to analyze the kernel core dump. As of now it contains the pointer to task struct of panicing task and page size. Page size is irrelevant for i386 but is required for architectures like ia64 and ppc64.

It's not absolutely required for crash -- but more of a convenience. As it is, the set of NT_PRSTATUS sections can be scoured for the set of active tasks, and then the process stacks (and potentially the IRQ stacks, then mapped back to the proper process stack) can be searched for crash_kexec. So crash works without the task_struct pointer, but it's ugly. It's just that netdump, diskdump and LKCD all report the panic task_struct pointer, and it seems a reasonable thing to do.

Quoted 11 lines
Quoted 2 lines

o gdb is not affected by this change as gdb need not to parse this note.

A couple of things.
- The name of your note is terribly generic, so it seems a poor choice.

- Why do we need to capture the page size at the time of the crash? Isn't the page size a compile time option? Won't sys_getpagesize() get you this information before the crash? Why do we need the kernels page size at all?

The page size is needed for a whole host of things, primarily for virtual-to-physical address translations, a bunch of VM-related commands, etc. For non-x86[_64] systems. the host system may be using a different page size than the vmcore being examining. That's probably a rare situation, and using getpagesize() on the host would work in most cases, or perhaps issuing the page size as a command line option, but again, it's another convenience item that gets reported by netdump, diskdump and LKCD.

Quoted 3 lines

- Why do you avoid storing the current task on the other cpus?

They can be found in the per-cpu runqueues data structure.

Quoted 9 lines

- Can't we derive the current task from the existing register information already captured. If not would a little extra debug information captured at compile time be better?

- You don't address the issue of architectural control registers. So you are going to need another note at some point. (Not necessarily a bad thing).

I guess that's their whole point. This NT_KDUMPINFO or whatever you want to call it, could cumulatively collect any other data deemed necessary that falls outside the bounds of the existing note types.

Dave Anderson

Quoted 20 lines
Quoted 20 lines
include/linux/elfcore.h renamed from include/linux/elfcore.h~kdump-x86-add-note-type-nt-kdump +13 -1
+/*
+ * NT_KDUMPINFO can be used to communicate additional information required for
+ * kernel core dumps. Additional information includes kernel configuration
+ * variables like page size and any other relevant data required by
+ * debugger (crash in this case).
+*/
+struct elf_kdumpinfo
+{
+     char page_shift;                /* Page size */
+     struct task_struct *panic_tsk;  /* Pointer to panic task_struct */
+};
+
 
 #define NT_TASKSTRUCT        4
 #define NT_AUXV              6
 #define NT_PRXFPREG 0x46e62b7f /* copied from gdb5.1/include/elf/common.h */
-
+#define NT_KDUMPINFO 7               /* Used for kernel core dumps */
#3 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Quoted 13 lines · Dave Anderson
Quoted 10 lines · Eric W. Biederman
Quoted 8 lines · Vivek Goyal

o This patch adds a new note type NT_KDUMPINFO. This note is added with kernel core dumps generated by kdump.

o This note mainly communicates the information required by kernel dump analysis tool "crash" to analyze the kernel core dump. As of now it contains the pointer to task struct of panicing task and page size. Page size is
irrelevant for i386 but is required for architectures like ia64 and ppc64.

Quoted 8 lines

It's not absolutely required for crash -- but more of a convenience. As it is, the
set of NT_PRSTATUS sections can be scoured for the set of active tasks, and then the process stacks (and potentially the IRQ stacks, then mapped back to the proper process stack) can be searched for crash_kexec. So crash works without the task_struct pointer, but it's ugly. It's just that netdump, diskdump and LKCD all report the panic task_struct pointer, and it seems a reasonable thing to do.

Ok. The point here is to know which task/cpu called panic, rather
than to get the task_struct. That makes a lot of sense, and is
cheap to get. Any note on the crashing cpu that is not captured by another cpu will give us that information.

My primary concern is while the concept of a task_struct is pretty stable who is to know how the kernel will change in the future. So if we don't need to export a task_struct pointer and merely need to flag the cpu that called panic we can do that much more reliably.

A practical question is how is this solved with a normal multi-threaded core dumps?

It is not the best thing but order of the notes does and will matter, it is likely worth Documenting so people don't change it that PR_STATUS comes first so we can reliably find the first note for a cpu, after everything has been serialized.

Quoted 11 lines
Quoted 10 lines
Quoted 1 line

o gdb is not affected by this change as gdb need not to parse this note.

A couple of things.
- The name of your note is terribly generic, so it seems a poor choice.

- Why do we need to capture the page size at the time of the crash? Isn't the page size a compile time option? Won't sys_getpagesize() get you this information before the crash? Why do we need the kernels page size at all?

Quoted 8 lines

The page size is needed for a whole host of things, primarily for virtual-to-physical address translations, a bunch of VM-related commands, etc. For non-x86[_64] systems. the host system may be using a different page size than the vmcore being examining. That's probably a rare situation, and using getpagesize() on the host would work in most cases, or perhaps issuing the page size as a command line option, but again, it's another convenience item that gets reported by netdump, diskdump and LKCD.

What I was mostly saying is that there are other places this can be captured. If this is actually a compile time option on all architectures we should simply add more debug information to vmlinux. If this varies at runtime we have the option of capturing it from the kernel before it crashes. We don't need to run code to capture it while the kernel is dying.

Having just skimmed the include/asm-* it looks like this is indeed a compile time option so the page size needs to get added to the information that is easy to get from the vmlinux binary.

We do need a way that we can test to see if a core dump actually matches the vmlinux we are looking at. Probably this is capturing all of the information captured by linux/version.h and linux/compile.h both at runtime and at compile time and checking to see if they match.

Quoted 4 lines
Quoted 2 lines

- Why do you avoid storing the current task on the other cpus?

They can be found in the per-cpu runqueues data structure.

Thanks. That confirms my impression we are just looking for a flag that says we are the current cpu.

Quoted 4 lines
Quoted 4 lines

- Can't we derive the current task from the existing register information already captured. If not would a little extra debug information captured at compile time be better?

Quoted 8 lines
Quoted 4 lines

- You don't address the issue of architectural control registers. So you are going to need another note at some point. (Not necessarily a bad thing).

I guess that's their whole point. This NT_KDUMPINFO or whatever you want to call it, could cumulatively collect any other data deemed necessary that falls outside the bounds of the existing note types.

It is clear we need one or mode notes. Largely once we define what is in a note we cannot change it. The most we can do is to append more fields to an existing node type, because we have size information we can check on. This is because the definition of a note is a kernel API issue.

Eric

#4 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Quoted 24 lines · Eric W. Biederman
Quoted 13 lines · Dave Anderson
Quoted 10 lines · Eric W. Biederman
Quoted 8 lines · Vivek Goyal

o This patch adds a new note type NT_KDUMPINFO. This note is added with kernel core dumps generated by kdump.

o This note mainly communicates the information required by kernel dump analysis tool "crash" to analyze the kernel core dump. As of now it contains the pointer to task struct of panicing task and page size. Page size is
irrelevant for i386 but is required for architectures like ia64 and ppc64.

Quoted 8 lines

It's not absolutely required for crash -- but more of a convenience. As it is, the
set of NT_PRSTATUS sections can be scoured for the set of active tasks, and then the process stacks (and potentially the IRQ stacks, then mapped back to the proper process stack) can be searched for crash_kexec. So crash works without the task_struct pointer, but it's ugly. It's just that netdump, diskdump and LKCD all report the panic task_struct pointer, and it seems a reasonable thing to do.

Quoted 5 lines

Ok. The point here is to know which task/cpu called panic, rather
than to get the task_struct. That makes a lot of sense, and is
cheap to get. Any note on the crashing cpu that is not captured by another cpu will give us that information.

That makes sense. Sheer presence of a particular note can identify the crashing cpu and no need to store "current". Only crashing cpu is going to store that note and that too after respective NT_PRSTATUS.

Quoted 44 lines

My primary concern is while the concept of a task_struct is pretty stable who is to know how the kernel will change in the future. So if we don't need to export a task_struct pointer and merely need to flag the cpu that called panic we can do that much more reliably.

A practical question is how is this solved with a normal multi-threaded core dumps?

It is not the best thing but order of the notes does and will matter, it is likely worth Documenting so people don't change it that PR_STATUS comes first so we can reliably find the first note for a cpu, after everything has been serialized.

Quoted 11 lines
Quoted 10 lines
Quoted 1 line

o gdb is not affected by this change as gdb need not to parse this note.

A couple of things.
- The name of your note is terribly generic, so it seems a poor choice.

- Why do we need to capture the page size at the time of the crash? Isn't the page size a compile time option? Won't sys_getpagesize() get you this information before the crash? Why do we need the kernels page size at all?

Quoted 8 lines

The page size is needed for a whole host of things, primarily for virtual-to-physical address translations, a bunch of VM-related commands, etc. For non-x86[_64] systems. the host system may be using a different page size than the vmcore being examining. That's probably a rare situation, and using getpagesize() on the host would work in most cases, or perhaps issuing the page size as a command line option, but again, it's another convenience item that gets reported by netdump, diskdump and LKCD.

What I was mostly saying is that there are other places this can be captured. If this is actually a compile time option on all architectures we should simply add more debug information to vmlinux. If this varies at runtime we have the option of capturing it from the kernel before it crashes. We don't need to run code to capture it while the kernel is dying.

Having just skimmed the include/asm-* it looks like this is indeed a compile time option so the page size needs to get added to the information that is easy to get from the vmlinux binary.

Quoted 6 lines

We do need a way that we can test to see if a core dump actually matches the vmlinux we are looking at. Probably this is capturing all of the information captured by linux/version.h and linux/compile.h both at runtime and at compile time and checking to see if they match.

I quickly browsed through "crash" code and looks like it is already doing similiar check (kernel.c, verify_version()). It seems to be retrieving "linux_banner" from core image and also retrieving banner string from vmlinux and trying to match. So if banner information can be directly read from the core image, probably there is no need to export it through notes.

Thanks
Vivek

#5 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Quoted 7 lines · Vivek Goyal
Quoted 6 lines

Ok. The point here is to know which task/cpu called panic, rather
than to get the task_struct. That makes a lot of sense, and is
cheap to get. Any note on the crashing cpu that is not captured by another cpu will give us that information.

Quoted 8 lines

That makes sense. Sheer presence of a particular note can identify the crashing cpu and no need to store "current". Only crashing cpu is going to store that note and that too after respective NT_PRSTATUS.

Quoted 4 lines

My primary concern is while the concept of a task_struct is pretty stable who is to know how the kernel will change in the future. So if we don't need to export a task_struct pointer and merely need to flag the cpu that called panic we can do that much more reliably.

Just flagging the cpu, and then mapping that to the stack pointer found in the associated NT_PRSTATUS register set should work OK too. It gets a little muddy if it crashed while running on an IRQ stack, but it still can be tracked back from there as well. (although not if the crashing task overflowed the IRQ stack)

The task_struct would be ideal though -- if the kernel's use of task_structs changes in the future, well, then crash is going to need a serious re-write anyway... FWIW, netdump and diskdump use the NT_TASKSTRUCT note note to store just the "current" pointer, and not the whole task_struct itself, which would just be a waste of space in the ELF header for crash's purposes. And looking at the gdb sources, it appears to be totally ignored. Who uses the NT_TASKSTRUCT note anyway?

Quoted 8 lines
Quoted 7 lines

We do need a way that we can test to see if a core dump actually matches the vmlinux we are looking at. Probably this is capturing all of the information captured by linux/version.h and linux/compile.h both at runtime and at compile time and checking to see if they match.

Quoted 6 lines

I quickly browsed through "crash" code and looks like it is already doing similiar check (kernel.c, verify_version()). It seems to be retrieving "linux_banner" from core image and also retrieving banner string from vmlinux and trying to match. So if banner information can be directly read from the core image, probably there is no need to export it through notes.

That's correct.

Dave

#6 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Quoted 5 lines · Dave Anderson

Just flagging the cpu, and then mapping that to the stack pointer found in the associated NT_PRSTATUS register set should work OK too. It gets a little muddy if it crashed while running on an IRQ stack, but it still can be tracked back from there as well. (although not if the crashing task overflowed the IRQ stack)

You can't track it back from the crashing cpu if the IRQ stack overflows either. So I would rather have crash confused when trying to find the task_struct. Then to have the kernel fail avoidably while attempting to capture a core dump.

Even if you overflow the stack wit a bit of detective work it should still be possible to show the stack overflowed and correct for it when analyzing the crash dump. Doing anything like that from a crashing cpu (in a reliable way) is very hard.

Quoted 7 lines

The task_struct would be ideal though -- if the kernel's use of task_structs changes in the future, well, then crash is going to need a serious re-write anyway... FWIW, netdump and diskdump use the NT_TASKSTRUCT note note to store just the "current" pointer, and not the whole task_struct itself, which would just be a waste of space in the ELF header for crash's purposes. And looking at the gdb sources, it appears to be totally ignored. Who uses the NT_TASKSTRUCT note anyway?

Good question, especially as the kernel exports whatever we have for a task struct today in the ELF note. No ABI compatibility is maintained.

Given all of that I recommend an empty NT_TASKSTRUCT to flag the crashing cpu, for now.

Eric

#7 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Sorry. Reposting since it did not made to LKML due to my stupid mistake; Also posting Dave Anderson's reponse:

Quoted 34 lines · Eric W. Biederman
Quoted 7 lines · Dave Anderson

Just flagging the cpu, and then mapping that to the stack pointer found in the associated NT_PRSTATUS register set should work OK too. It gets a little muddy if it crashed while running on an IRQ stack, but it still can be tracked back from there as well. (although not if the crashing task overflowed the IRQ stack)

You can't track it back from the crashing cpu if the IRQ stack overflows either. So I would rather have crash confused when trying to find the task_struct. Then to have the kernel fail avoidably while attempting to capture a core dump.

Even if you overflow the stack wit a bit of detective work it should still be possible to show the stack overflowed and correct for it when analyzing the crash dump. Doing anything like that from a crashing cpu (in a reliable way) is very hard.

Quoted 9 lines

The task_struct would be ideal though -- if the kernel's use of task_structs changes in the future, well, then crash is going to need a serious re-write anyway... FWIW, netdump and diskdump use the NT_TASKSTRUCT note note to store just the "current" pointer, and not the whole task_struct itself, which would just be a waste of space in the ELF header for crash's purposes. And looking at the gdb sources, it appears to be totally ignored. Who uses the NT_TASKSTRUCT note anyway?

Quoted 8 lines

Good question, especially as the kernel exports whatever we have for a task struct today in the ELF note. No ABI compatibility is maintained.

Given all of that I recommend an empty NT_TASKSTRUCT to flag the crashing cpu, for now.

At present /proc/kcore writes the complete task structure for NT_TASKSTRUCT note section. Thought it is the standard. Hence created separate note section. The other option is the crash tool can directly read "crashing_cpu variable" from the vmcore to determine the panic cpu. Similarly, we can define panic_task variable in the kernel.

Dave Anderson ([email protected]) reponse:

" So does elf_core_dump() as well, but to gdb it's useless AFAICT...

Hey -- I wasn't even aware of the "crashing_cpu" variable. That would work just fine.

Still a "panic_task", and perhaps even a "crash_page_size" variable
would be nice as well. No additional notes required...

Dave "

Basically, we can use some global structure in the kernel and dump any needed information which we do not need to invoke any analysis tools (crash, gdb). Dumping CPU control registers can also be done this way without creating separate note section.

Thanks
Haren

Quoted 11 lines

Eric

------------------------------------------------------------------------

_______________________________________________
fastboot mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/fastboot

#8 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

In reply to Eric W. Biederman
Quoted 26 lines · Eric W. Biederman
Quoted 5 lines · Dave Anderson

Just flagging the cpu, and then mapping that to the stack pointer found in the associated NT_PRSTATUS register set should work OK too. It gets a little muddy if it crashed while running on an IRQ stack, but it still can be tracked back from there as well. (although not if the crashing task overflowed the IRQ stack)

You can't track it back from the crashing cpu if the IRQ stack overflows either. So I would rather have crash confused when trying to find the task_struct. Then to have the kernel fail avoidably while attempting to capture a core dump.

Even if you overflow the stack wit a bit of detective work it should still be possible to show the stack overflowed and correct for it when analyzing the crash dump. Doing anything like that from a crashing cpu (in a reliable way) is very hard.

Quoted 7 lines

The task_struct would be ideal though -- if the kernel's use of task_structs changes in the future, well, then crash is going to need a serious re-write anyway... FWIW, netdump and diskdump use the NT_TASKSTRUCT note note to store just the "current" pointer, and not the whole task_struct itself, which would just be a waste of space in the ELF header for crash's purposes. And looking at the gdb sources, it appears to be totally ignored. Who uses the NT_TASKSTRUCT note anyway?

Quoted 7 lines

Good question, especially as the kernel exports whatever we have for a task struct today in the ELF note. No ABI compatibility is maintained.

Given all of that I recommend an empty NT_TASKSTRUCT to flag the crashing cpu, for now.

I got a concern here. Are we not breaking the convention. NT_TASKSTRUCT note type represents that task_sturct is stored in note data. elf_core_dump() and /proc/kcore already do that (That's a different story that it might not be needed at all). Now if we store a null NT_TASKSTURCT, same note type will carry two meanings.

IMHO, introducing a null NT_KDUMPINFO will help in that sense, at least there are no two interpretations of same note type. Also it provides the scope to add more elements to it if need be.

Thanks
Vivek

#9 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Quoted 5 lines · Vivek Goyal

I got a concern here. Are we not breaking the convention. NT_TASKSTRUCT note type represents that task_sturct is stored in note data. elf_core_dump() and /proc/kcore already do that (That's a different story that it might not be needed at all). Now if we store a null NT_TASKSTURCT, same note type will carry two meanings.

My impression is that NT_TASKSTRUCT has failed to define what actually appears in the note, as the kernel task_struct is not designed to be exported to user space, and can change without warning.

Hmm. NT_TASKSTRUCT feels like an information leak although I haven't a clue what kernel information you could leak that would be interesting enough that you could compromise the kernel. It looks like we should export NT_TASKSTRUCT until we can decide what goes there.

Quoted 3 lines

IMHO, introducing a null NT_KDUMPINFO will help in that sense, at least there are no two interpretations of same note type. Also it provides the scope to add more elements to it if need be.

Agreed. I was trying too hard to reuse what was already there.

Eric

#10 lore raw

Subject: [PATCH] Don't uselessly export task_struct to user space in core dumps.

In reply to Vivek Goyal

task_struct is an internal structure to the kernel with a lot of good information, that is probably interesting in core dumps. However there is no way for user space to know what format that information is in making it useless.

I grepped the GDB 6.3 source code and NT_TASKSTRUCT while defined is not used anywhere else. So I would be surprised if anyone notices it is missing.

In addition exporting kernel pointers to all the interesting kernel data structures sounds like the very definition of an information leak. I haven't a clue what someone with evil intentions could do with that information, but in any attack against the kernel it looks like this is the perfect tool for aiming that attack.

So since NT_TASKSTRUCT is useless as currently defined and is potentially dangerous, let's just not export it.

---

 arch/mips/kernel/irixelf.c |   17 ++++++-----------
 fs/binfmt_elf.c            |    4 +---
 2 files changed, 7 insertions(+), 14 deletions(-)

06d116247caa2fef10cb73c5d6042a5ff658b677

arch/mips/kernel/irixelf.c +6 -11
@@ -1064,8 +1064,8 @@ static int irix_core_dump(long signr, st
 	struct elfhdr elf;
 	off_t offset = 0, dataoff;
 	int limit = current->signal->rlim[RLIMIT_CORE].rlim_cur;
-	int numnote = 4;
-	struct memelfnote notes[4];
+	int numnote = 3;
+	struct memelfnote notes[3];
 	struct elf_prstatus prstatus;	/* NT_PRSTATUS */
 	elf_fpregset_t fpu;		/* NT_PRFPREG */
 	struct elf_prpsinfo psinfo;	/* NT_PRPSINFO */
@@ -1198,20 +1198,15 @@ static int irix_core_dump(long signr, st
 	}
 	strlcpy(psinfo.pr_fname, current->comm, sizeof(psinfo.pr_fname));
 
-	notes[2].name = "CORE";
-	notes[2].type = NT_TASKSTRUCT;
-	notes[2].datasz = sizeof(*current);
-	notes[2].data = current;
-
 	/* Try to dump the FPU. */
 	prstatus.pr_fpvalid = dump_fpu (regs, &fpu);
 	if (!prstatus.pr_fpvalid) {
 		numnote--;
 	} else {
-		notes[3].name = "CORE";
-		notes[3].type = NT_PRFPREG;
-		notes[3].datasz = sizeof(fpu);
-		notes[3].data = &fpu;
+		notes[2].name = "CORE";
+		notes[2].type = NT_PRFPREG;
+		notes[2].datasz = sizeof(fpu);
+		notes[2].data = &fpu;
 	}
 
 	/* Write notes phdr entry. */
fs/binfmt_elf.c +1 -3
@@ -1503,9 +1503,7 @@ static int elf_core_dump(long signr, str
 	fill_psinfo(psinfo, current->group_leader, current->mm);
 	fill_note(notes +1, "CORE", NT_PRPSINFO, sizeof(*psinfo), psinfo);
 	
-	fill_note(notes +2, "CORE", NT_TASKSTRUCT, sizeof(*current), current);
-  
-	numnote = 3;
+	numnote = 2;
 
 	auxv = (elf_addr_t *) current->mm->saved_auxv;
 

A message that is not in the archive (sent off-list or not yet received): OF0A1E6B6F.F00DC760-ON87257084.005F99D6-88257084.00634A38@us.ibm.com

#12 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

Quoted 41 lines · Haren Myneni

[email protected] wrote on 09/22/2005 09:31:52 AM:

Quoted 3 lines
Quoted 1 line · Dave Anderson

Just flagging the cpu, and then mapping that to the stack pointer

found in

Quoted 5 lines
Quoted 2 lines

the associated NT_PRSTATUS register set should work OK too. It gets a little muddy if it crashed while running on an IRQ stack, but it

still can be

Quoted 1 line

tracked back from there as well. (although not if the crashing

task overflowed

Quoted 8 lines
Quoted 1 line

the IRQ stack)

You can't track it back from the crashing cpu if the IRQ stack overflows either. So I would rather have crash confused when trying to find the task_struct. Then to have the kernel fail avoidably while attempting to capture a core dump.

Even if you overflow the stack wit a bit of detective work it should

still

Quoted 1 line

be possible to show the stack overflowed and correct for it when

analyzing

Quoted 4 lines

the crash dump. Doing anything like that from a crashing cpu (in a reliable way) is very hard.

Quoted 1 line

The task_struct would be ideal though -- if the kernel's use of

task_structs

Quoted 1 line
Quoted 1 line

changes in the future, well, then crash is going to need a serious

re-write

Quoted 4 lines
Quoted 2 lines

anyway... FWIW, netdump and diskdump use the NT_TASKSTRUCT note note to store just the "current" pointer, and not the whole

task_struct itself,

Quoted 1 line

which would just be a waste of space in the ELF header for

crash'spurposes.

Quoted 7 lines
Quoted 2 lines

And looking at the gdb sources, it appears to be totally ignored. Who uses the NT_TASKSTRUCT note anyway?

Good question, especially as the kernel exports whatever we have for a task struct today in the ELF note. No ABI compatibility is maintained.

Quoted 8 lines
Quoted 2 lines

Given all of that I recommend an empty NT_TASKSTRUCT to flag the crashing cpu, for now.

At present /proc/kcore writes the complete task structure for NT_TASKSTRUCT note section. Thought it is the standard. Hence created separate note section. The other option is the crash tool can directly read "crashing_cpu variable" from the vmcore to determine the panic cpu. Similarly, we can define panic_task variable in the kernel.

crashing_cpu was introduced recently to handle one of the problems caused due to NMI. I think we should not be relying on this variable. we get the value of crashing_cpu by making smp_processor_id() call and this value will be corrupted in case of stack overflow.

During OLS, Eric had suggested to either find a way to disable NMI after panic() or may be read LAPIC id. Reading LAPIC id seems to be more reliable. I think mkdump folks already do something similar.

So, in short, using crashing_cpu might not be a good idea. Down the line this varibale might not be present at all.

Quoted 21 lines

Basically, we can use some global structure in the kernel and dump any needed information which we do not need to invoke any analysis tools (crash, gdb). Dumping CPU control registers can also be done this way without creating separate note section.

Thanks
Haren

Anyway, we already have crashing_cpu variable in the kernel.

Quoted 6 lines

Eric
_______________________________________________
fastboot mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/fastboot

_______________________________________________
fastboot mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/fastboot

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

#14 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernelcore dumps

Quoted 1 line · Dave Anderson

So does elf_core_dump() as well, but to gdb it's useless AFAICT...

We can always post_process things when generating a core dump if we have enough information.

Quoted 5 lines

Hey -- I wasn't even aware of the "crashing_cpu" variable. That would work just fine.

Still a "panic_task", and perhaps even a "crash_page_size" variable
would be nice as well. No additional notes required...

To avoid defining an ABI that we need to maintain there is some benefit in simply using static variables. But the form of the information really isn't the concern.

Where we capture the information and how reliable is that capture is the concern.

To capture page size the easiest and most reliable way I can see to do is to modify vmlinux.lds.S to contain something like:

Quoted 1 line
	 _page_shift = PAGE_SHIFT;

Giving you an absolute symbol _page_shift in vmlinux that contains the value you need, without overhead in the running kernel.

crashing_cpu makes sense to capture in some form, we definitely need to compute something that will allow us to write to a per cpu area on an SMP system.

The big concern at this point is that the code has not undergone a serious stability audit. So it is the expectation that there is still code we can remove and modify to increase the likely hood of getting a crash dump.

Currently we know that stack overflows sometimes happen and that they are a source of kernel crashes. It would be good if we could take a crash dump despite them. To do that requires code more robust than we have today. Quite likely it means that we will not be able to reliably capture the task_struct of the crashing cpu.

Eric

#15 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernelcore dumps

Quoted 20 lines · Eric W. Biederman
Quoted 1 line · Dave Anderson

So does elf_core_dump() as well, but to gdb it's useless AFAICT...

We can always post_process things when generating a core dump if we have enough information.

Quoted 5 lines

Hey -- I wasn't even aware of the "crashing_cpu" variable. That would work just fine.

Still a "panic_task", and perhaps even a "crash_page_size" variable
would be nice as well. No additional notes required...

To avoid defining an ABI that we need to maintain there is some benefit in simply using static variables. But the form of the information really isn't the concern.

Where we capture the information and how reliable is that capture is the concern.

Quoted 6 lines

To capture page size the easiest and most reliable way I can see to do is to modify vmlinux.lds.S to contain something like:

Quoted 1 line
	 _page_shift = PAGE_SHIFT;

Giving you an absolute symbol _page_shift in vmlinux that contains the value you need, without overhead in the running kernel.

That's a cool idea. I just tested it.

[vivek@vivegoya linux]$ readelf -s vmlinux | grep __page_shift
28865: 0000000c 0 NOTYPE GLOBAL DEFAULT ABS __page_shift

Quoted 10 lines

crashing_cpu makes sense to capture in some form, we definitely need to compute something that will allow us to write to a per cpu area on an SMP system.

The big concern at this point is that the code has not undergone a serious stability audit. So it is the expectation that there is still code we can remove and modify to increase the likely hood of getting a crash dump.

Quoted 6 lines

Currently we know that stack overflows sometimes happen and that they are a source of kernel crashes. It would be good if we could take a crash dump despite them. To do that requires code more robust than we have today. Quite likely it means that we will not be able to reliably capture the task_struct of the crashing cpu.

Agreed.

#16 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernelcore dumps

In reply to Eric W. Biederman
Quoted 36 lines · Eric W. Biederman
Quoted 1 line · Dave Anderson

So does elf_core_dump() as well, but to gdb it's useless AFAICT...

We can always post_process things when generating a core dump if we have enough information.

Quoted 5 lines

Hey -- I wasn't even aware of the "crashing_cpu" variable. That would work just fine.

Still a "panic_task", and perhaps even a "crash_page_size" variable
would be nice as well. No additional notes required...

To avoid defining an ABI that we need to maintain there is some benefit in simply using static variables. But the form of the information really isn't the concern.

Where we capture the information and how reliable is that capture is the concern.

To capture page size the easiest and most reliable way I can see to do is to modify vmlinux.lds.S to contain something like:

Quoted 1 line
	 _page_shift = PAGE_SHIFT;

Giving you an absolute symbol _page_shift in vmlinux that contains the value you need, without overhead in the running kernel.

crashing_cpu makes sense to capture in some form, we definitely need to compute something that will allow us to write to a per cpu area on an SMP system.

The big concern at this point is that the code has not undergone a serious stability audit. So it is the expectation that there is still code we can remove and modify to increase the likely hood of getting a crash dump.

Quoted 6 lines

Currently we know that stack overflows sometimes happen and that they are a source of kernel crashes. It would be good if we could take a crash dump despite them. To do that requires code more robust than we have today. Quite likely it means that we will not be able to reliably capture the task_struct of the crashing cpu.

Ok. After all this discussion looks like time to drop the patch. The architectures which need page_size, they can modify vmlinux.ld.S to embed the information in vmlinux.

Crash can do without pointer to crashing cpu's task_struct as well (Though it is more work on crash's part). At the same time this information can not possibly be captured in kernel reliably after the crash.

Identifying crashing cpu is important. Empty NT_KDUMPINFO would have done the job but use of variables will avoid defining ABI. For the time being crashing_cpu will do the job. Ofcourse this code is not robust enough to handle stack overflow situations and that shall have to be taken care of.

Andrew, Could you please drop this patch from -mm.

kdumpx86-add-note-type-nt_kdumpinfo-to-kernel-core-dumps.patch

Thanks
Vivek

Quoted 6 lines

Eric

_______________________________________________
fastboot mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/fastboot

#17 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFO tokernel core dumps

In reply to Vivek Goyal
Quoted 5 lines · Vivek Goyal

I quickly browsed through "crash" code and looks like it is already doing similiar check (kernel.c, verify_version()). It seems to be retrieving "linux_banner" from core image and also retrieving banner string from vmlinux and trying to match. So if banner information can be directly read from the core image, probably there is no need to export it through notes.

Sounds good. We still need to define a note for the cpu control registers. Do any of the other crash dump solution capture that information right now?

Eric

#18 lore raw

Re: [Fastboot] [PATCH] Kdump(x86): add note type NT_KDUMPINFOtokernel core dumps

Quoted 8 lines · Eric W. Biederman
Quoted 5 lines · Vivek Goyal

I quickly browsed through "crash" code and looks like it is already doing similiar check (kernel.c, verify_version()). It seems to be retrieving "linux_banner" from core image and also retrieving banner string from vmlinux and trying to match. So if banner information can be directly read from the core image, probably there is no need to export it through notes.

Quoted 5 lines

Sounds good. We still need to define a note for the cpu control registers. Do any of the other crash dump solution capture that information right now?

Eric

Certainly not in netdump, diskdump or LKCD...

On the other hand, I can't say I ever really needed it, although that's not to say it couldn't be valuable for some types of crashes.

Dave

#19 lore raw
In reply to Eric W. Biederman
Quoted 10 lines · Eric W. Biederman
Quoted 8 lines · Vivek Goyal

o This patch adds a new note type NT_KDUMPINFO. This note is added with kernel core dumps generated by kdump.

o This note mainly communicates the information required by kernel dump analysis tool "crash" to analyze the kernel core dump. As of now it contains the pointer to task struct of panicing task and page size. Page size is irrelevant for i386 but is required for architectures like ia64 and ppc64.

Quoted 5 lines
Quoted 1 line

o gdb is not affected by this change as gdb need not to parse this note.

A couple of things.
- The name of your note is terribly generic, so it seems a poor choice.

Yes it seems generic. And I think the reason being that I did not have a significant number of things to report hence could not create logical groups and giving very specific names to those groups. As of today, based on the requirements, could think of only two items (page size, current). There is still scope for few more things to appear. Hence gave a generic name and thought more items can be grouped in this note itself until and unless group becomes really big and need to break.

Do you have any suggestion to make this name better?

Quoted 5 lines

- Why do we need to capture the page size at the time of the crash? Isn't the page size a compile time option? Won't sys_getpagesize() get you this information before the crash? Why do we need the kernels page size at all?

Dave has already mentioned about the need of page size. I agree that page size is a compile time information. Are you hinting at creating a separate note for reporting page size in user space while loading the capture kernel and store it in reserved region alongside other elf headers. Not sure, if it is good to create a separate note type just to report one configuration variable.

I also thought of enabling "Kernel .config support" (CONFIG_IKCONFIG=y) and may be "crash" could write a script similar to scripts/extract-ikconfig to retrieve any configuration variable required from vmlinux itself. But this option is not enabled by default and will increase kernel image size and seems too much to do for a single configuration variable.

Quoted 4 lines

- Why do you avoid storing the current task on the other cpus?

- Can't we derive the current task from the existing register information already captured.

It can be done but as Dave suggested but that requires significant amount of job to be done as one has to traverse through the active task stacks and look for crash_kexec(). An easier/simpler way is that kernel itself can report it. Netdump, diskdump already do it. I think for simplicity, it makes sense to export this information from kernel in the form of note.

Only issue I could think of is stack overflow and current might be corrupted after panic.

Quoted 3 lines

If not would a little extra debug information
captured at compile time be better?

Sorry, I did not get this. Can you elaborate a little more on this. What kind of debug information can be helpful in determining the current.

Quoted 4 lines

- You don't address the issue of architectural control registers. So you are going to need another note at some point. (Not necessarily a bad thing).

That's true. Probably a new note type shall be required to store architectural control registers. May be something like NT_KDUMP_CRREG and NT_KDUMPINFO can continue to capture miscellaneous info.

Thanks
Vivek

Quoted 20 lines
Quoted 20 lines
include/linux/elfcore.h renamed from include/linux/elfcore.h~kdump-x86-add-note-type-nt-kdump +13 -1
+/*
+ * NT_KDUMPINFO can be used to communicate additional information required for
+ * kernel core dumps. Additional information includes kernel configuration
+ * variables like page size and any other relevant data required by
+ * debugger (crash in this case).
+*/
+struct elf_kdumpinfo
+{
+	char page_shift;		/* Page size */
+	struct task_struct *panic_tsk;	/* Pointer to panic task_struct */
+};
+
 
 #define NT_TASKSTRUCT	4
 #define NT_AUXV		6
 #define NT_PRXFPREG 0x46e62b7f /* copied from gdb5.1/include/elf/common.h */
-
+#define NT_KDUMPINFO	7		/* Used for kernel core dumps */
#20 lore raw
Quoted 6 lines · Vivek Goyal
Quoted 5 lines

- Why do you avoid storing the current task on the other cpus?

- Can't we derive the current task from the existing register information already captured.

Quoted 9 lines

It can be done but as Dave suggested but that requires significant amount of job to be done as one has to traverse through the active task stacks and look for crash_kexec(). An easier/simpler way is that kernel itself can report it. Netdump, diskdump already do it. I think for simplicity, it makes sense to export this information from kernel in the form of note.

Only issue I could think of is stack overflow and current might be corrupted after panic.

Yes, traversing the task_structs in a crashed kernel sounds like a poor idea.

#21 lore raw
Quoted 13 lines · Andrew Morton
Quoted 12 lines · Vivek Goyal
Quoted 5 lines

- Why do you avoid storing the current task on the other cpus?

- Can't we derive the current task from the existing register information already captured.

It can be done but as Dave suggested but that requires significant amount of job to be done as one has to traverse through the active task stacks and look for crash_kexec(). An easier/simpler way is that kernel itself can report it. Netdump, diskdump already do it. I think for simplicity, it makes sense to export this information from kernel in the form of note.

Quoted 7 lines
Quoted 3 lines

Only issue I could think of is stack overflow and current might be corrupted after panic.

Yes, traversing the task_structs in a crashed kernel sounds like a poor idea.

Andrew, traversal of task_structs will not be done in crashed kernel. It will be done in user space by "crash" (if required, during post crash analysis).

In this case we are simply copying "current" pointer of panicking task in an elf note.

	kdumpinfo.panic_tsk = current;

Thanks
Vivek

#22 lore raw
In reply to Vivek Goyal
Quoted 17 lines · Vivek Goyal
Quoted 10 lines · Eric W. Biederman
Quoted 8 lines · Vivek Goyal

o This patch adds a new note type NT_KDUMPINFO. This note is added with kernel core dumps generated by kdump.

o This note mainly communicates the information required by kernel dump analysis tool "crash" to analyze the kernel core dump. As of now it contains the pointer to task struct of panicing task and page size. Page size is
irrelevant for i386 but is required for architectures like ia64 and ppc64.

Quoted 5 lines
Quoted 1 line

o gdb is not affected by this change as gdb need not to parse this note.

A couple of things.
- The name of your note is terribly generic, so it seems a poor choice.

Quoted 7 lines

Yes it seems generic. And I think the reason being that I did not have a significant number of things to report hence could not create logical groups and giving very specific names to those groups. As of today, based on the requirements, could think of only two items (page size, current). There is still scope for few more things to appear. Hence gave a generic name and thought more items can be grouped in this note itself until and unless group becomes really big and need to break.

I'm not certain exactly how but we do need to capture the kernel version so we can verify the vmlinux binary and the core dump match.

Quoted 1 line

Do you have any suggestion to make this name better?

Not presently. That still doesn't mean it isn't worth brain storming about.

Quoted 6 lines
Quoted 5 lines

- Why do we need to capture the page size at the time of the crash? Isn't the page size a compile time option? Won't sys_getpagesize() get you this information before the crash? Why do we need the kernels page size at all?

Quoted 6 lines

Dave has already mentioned about the need of page size. I agree that page size is a compile time information. Are you hinting at creating a separate note for reporting page size in user space while loading the capture kernel and store it in reserved region alongside other elf headers. Not sure, if it is good to create a separate note type just to report one configuration variable.

It's in my reply to Dave but we need a way to get that kind of information out of vmlinux.

Quoted 4 lines
Quoted 4 lines

- Why do you avoid storing the current task on the other cpus?

- Can't we derive the current task from the existing register information already captured.

In my reply to Dave I noted that flagging the cpu is cheaper and more reliably that returning current.

Quoted 6 lines
Quoted 3 lines

If not would a little extra debug information
captured at compile time be better?

Sorry, I did not get this. Can you elaborate a little more on this. What kind of debug information can be helpful in determining the current.

The dwarf2 debug information can tell you what registers a variable is stored in. I was thinking of something along those lines, for reporting to the debugger where current was stored.

Quoted 8 lines
Quoted 4 lines

- You don't address the issue of architectural control registers. So you are going to need another note at some point. (Not necessarily a bad thing).

That's true. Probably a new note type shall be required to store architectural control registers. May be something like NT_KDUMP_CRREG and NT_KDUMPINFO can continue to capture miscellaneous info.

Sounds right. I am slightly less concerned so long as we restrict ourselves to an append only discipline with regards to the information in that note.

Eric