12:38 a.m.
There are 3 bugs, which are caused due to newer kernel change, they were
found when testing the gdb stack unwinding support[1] for ppc64 against the
6.x kernels. See the discussion in [2][3]. So they'd better to be fixed
in a batch.
The 3 bugs are:
1) STACK_SWITCH_FRAME_REGS replaced STACK_FRAME_OVERHEAD for determine
the register values offset within stack. Fixed in patch 1/2.
2) EXCP_FRAME_MARKER outdated for determine the eframe stack. Fixed in
patch 2/2.
3) inline functions not inlined in kernel. Should be fixed in kernel
side rather than crash, unhandled currently.
By the way, currently the gdb stack unwinding support requires the
following patchsets to be merged first:
1. This patchset;
2. [Crash-utility] [PATCH v2] Fix the regression of cpumask_t for xen
hyper;
3. Patch for the No.3 bug, if we want to test against the 6.X el10 kernel
against gdb stack unwinding. Since the patch is not ready, 5.X el9 kernel
can be tested instead.
[1]: https://www.mail-archive.com/devel@lists.crash-utility.osci.io/msg01079.html
[2]: https://www.mail-archive.com/devel@lists.crash-utility.osci.io/msg01116.html
[3]: https://www.mail-archive.com/devel@lists.crash-utility.osci.io/msg01126.html
v1 -> v2:
1) Get rid of instruction disassembly to determine if abi_v2 enabled.
2) Updated kernel commit hash into the crash patches.
Tao Liu (2):
ppc64: Fix bt printing error stack trace
ppc64: fix the bug eframe won't print for newer kernel
defs.h | 9 +++++--
netdump.c | 14 +++++++----
ppc64.c | 71 +++++++++++++++++++++++++++++++++++++++++++++----------
symbols.c | 5 ++--
4 files changed, 78 insertions(+), 21 deletions(-)
--
2.46.2
12:38 a.m.
New subject: [PATCH v2 1/2] ppc64: Fix bt printing error stack trace
A error stack trace of bt cmd observed:
crash> bt 1
PID: 1 TASK: c000000003714b80 CPU: 2 COMMAND: "systemd"
#0 [c0000000037735c0] _end at c0000000037154b0 (unreliable)
#1 [c000000003773770] __switch_to at c00000000001fa9c
#2 [c0000000037737d0] __schedule at c00000000112e4ec
#3 [c0000000037738b0] schedule at c00000000112ea80
...
The #0 stack trace is incorrect, the function address shouldn't exceed _end.
The reason is for kernel commit cd52414d5a6c ("powerpc/64: ELFv2 use
minimal stack frames in int and switch frame sizes"), the offset of pt_regs
to sp changed from STACK_FRAME_OVERHEAD, i.e 112, to STACK_SWITCH_FRAME_REGS.
For CONFIG_PPC64_ELF_ABI_V1, it's 112, for ABI_V2, it's 48. So the nip will
read a wrong value from stack when ABI_V2 enabled.
After the patch:
crash> bt 1
PID: 1 TASK: c000000003714b80 CPU: 2 COMMAND: "systemd"
#0 [c0000000037737d0] __schedule at c00000000112e4ec
#1 [c0000000037738b0] schedule at c00000000112ea80
...
Signed-off-by: Tao Liu <ltao(a)redhat.com>
Suggested-by: Aditya Gupta <adityag(a)linux.ibm.com>
---
defs.h | 3 ++-
netdump.c | 14 ++++++++++----
ppc64.c | 34 +++++++++++++++++++++++++++++++---
symbols.c | 5 +++--
4 files changed, 46 insertions(+), 10 deletions(-)
diff --git a/defs.h b/defs.h
index e2a9278..abde47c 100644
--- a/defs.h
+++ b/defs.h
@@ -4650,6 +4650,7 @@ struct efi_memory_desc_t {
#define MSR_PR_LG 14 /* Problem State / Privilege Level */
/* Used to find the user or kernel-mode frame*/
+#define STACK_SWITCH_FRAME_REGS 48
#define STACK_FRAME_OVERHEAD 112
#define EXCP_FRAME_MARKER 0x7265677368657265
@@ -5770,7 +5771,7 @@ void dump_offset_table(char *, ulong);
int is_elf_file(char *);
int is_kernel(char *);
int is_shared_object(char *);
-int file_elf_version(char *);
+int file_elf_header(char *, char *);
int is_system_map(char *);
int is_compressed_kernel(char *, char **);
int select_namelist(char *);
diff --git a/netdump.c b/netdump.c
index b4e2a5c..8fdc530 100644
--- a/netdump.c
+++ b/netdump.c
@@ -646,11 +646,11 @@ resize_elf_header(int fd, char *file, char **eheader_ptr, char
**sect0_ptr,
}
/*
- * Return the e_version number of an ELF file
+ * Return the e_version & e_flags number of an ELF file
* (or -1 if its not readable ELF file)
*/
int
-file_elf_version(char *file)
+file_elf_header(char *file, char *member)
{
int fd, size;
Elf32_Ehdr *elf32;
@@ -680,11 +680,17 @@ file_elf_version(char *file)
(elf32->e_ident[EI_CLASS] == ELFCLASS32) &&
(elf32->e_ident[EI_DATA] == ELFDATA2LSB) &&
(elf32->e_ident[EI_VERSION] == EV_CURRENT)) {
- return (elf32->e_version);
+ if (STRNEQ(member, "e_version"))
+ return (elf32->e_version);
+ else if (STRNEQ(member, "e_flags"))
+ return (elf32->e_flags);
} else if (STRNEQ(elf64->e_ident, ELFMAG) &&
(elf64->e_ident[EI_CLASS] == ELFCLASS64) &&
(elf64->e_ident[EI_VERSION] == EV_CURRENT)) {
- return (elf64->e_version);
+ if (STRNEQ(member, "e_version"))
+ return (elf64->e_version);
+ else if (STRNEQ(member, "e_flags"))
+ return (elf64->e_flags);
}
return -1;
diff --git a/ppc64.c b/ppc64.c
index e8930a1..87f9c4c 100644
--- a/ppc64.c
+++ b/ppc64.c
@@ -72,6 +72,7 @@ static ulong pud_page_vaddr_l4(ulong pud);
static ulong pmd_page_vaddr_l4(ulong pmd);
static int is_opal_context(ulong sp, ulong nip);
void opalmsg(void);
+static bool is_ppc64_elf_abi_v2(void);
static int is_opal_context(ulong sp, ulong nip)
{
@@ -2813,6 +2814,25 @@ ppc64_get_sp(ulong task)
return sp;
}
+static bool
+is_ppc64_elf_abi_v2(void)
+{
+ static bool ret = false;
+ static bool checked = false;
+
+ if (checked)
+ return ret;
+ switch (file_elf_header(pc->namelist, "e_flags")) {
+ case 2:
+ ret = true;
+ case 1:
+ break;
+ default:
+ error(WARNING, "Unstable elf_abi v1/v2 detection.\n");
+ }
+ checked = true;
+ return ret;
+}
/*
* get the SP and PC values for idle tasks.
@@ -2834,9 +2854,17 @@ get_ppc64_frame(struct bt_info *bt, ulong *getpc, ulong *getsp)
sp = ppc64_get_sp(task);
if (!INSTACK(sp, bt))
goto out;
- readmem(sp+STACK_FRAME_OVERHEAD, KVADDR, ®s,
- sizeof(struct ppc64_pt_regs),
- "PPC64 pt_regs", FAULT_ON_ERROR);
+
+ if (THIS_KERNEL_VERSION >= LINUX(6,2,0) && is_ppc64_elf_abi_v2()) {
+ readmem(sp+STACK_SWITCH_FRAME_REGS, KVADDR, ®s,
+ sizeof(struct ppc64_pt_regs),
+ "PPC64 pt_regs", FAULT_ON_ERROR);
+ } else {
+ readmem(sp+STACK_FRAME_OVERHEAD, KVADDR, ®s,
+ sizeof(struct ppc64_pt_regs),
+ "PPC64 pt_regs", FAULT_ON_ERROR);
+ }
+
ip = regs.nip;
closest = closest_symbol(ip);
if (STREQ(closest, ".__switch_to") || STREQ(closest, "__switch_to"))
{
diff --git a/symbols.c b/symbols.c
index d00fbd7..840e455 100644
--- a/symbols.c
+++ b/symbols.c
@@ -216,7 +216,7 @@ symtab_init(void)
* Check whether the namelist is a kerntypes file built by
* dwarfextract, which places a magic number in e_version.
*/
- if (file_elf_version(pc->namelist) == EV_DWARFEXTRACT)
+ if (file_elf_header(pc->namelist, "e_version") == EV_DWARFEXTRACT)
pc->flags |= KERNTYPES;
if (pc->flags & SYSMAP) {
@@ -13005,7 +13005,8 @@ load_module_symbols(char *modref, char *namelist, ulong
base_addr)
error(FATAL, "cannot determine object file format: %s\n",
namelist);
- if (LKCD_KERNTYPES() && (file_elf_version(namelist) == EV_DWARFEXTRACT))
+ if (LKCD_KERNTYPES() &&
+ (file_elf_header(namelist, "e_version") == EV_DWARFEXTRACT))
goto add_symbols; /* no symbols, add the debuginfo */
if (!(bfd_get_file_flags(mbfd) & HAS_SYMS))
--
2.46.2
12:38 a.m.
New subject: [PATCH v2 2/2] ppc64: fix the bug eframe won't print for newer kernel
The value of EXCP_FRAME_MARKER is outdated for kernel>=6.1.0 by kernel
commit bbd71709087a9 ("powerpc: Make stack frame marker upper case") and
commit 19c95df1277c4 ("powerpc: Reverse stack frame marker on little endian").
Also its offset is changed since kernel>=6.2.0 by kernel commit
cd52414d5a6c ("powerpc/64: ELFv2 use minimal stack frames in int and switch
frame sizes"). This patch will fix the issue by updating its values.
Before:
crash> bt
...
#8 [c00000000a327b80] system_call_exception at c000000000031a30
#9 [c00000000a327e50] system_call_vectored_common at c00000000000d05c
After:
crash> bt
...
#8 [c00000000a327b80] system_call_exception at c000000000031a30
#9 [c00000000a327e50] system_call_vectored_common at c00000000000d05c
System Call Vectored [3000] exception frame:
R0: 0000000000000004 R1: 00007fffc3345f40 R2: 0000000000100000
R3: 0000000000000001 R4: 000000012572c620 R5: 0000000000000002
R6: 0000000000000001 R7: 0000000000000004 R8: 0000000000000001
...
Signed-off-by: Tao Liu <ltao(a)redhat.com>
---
defs.h | 6 +++++-
ppc64.c | 37 +++++++++++++++++++++++++++----------
2 files changed, 32 insertions(+), 11 deletions(-)
diff --git a/defs.h b/defs.h
index abde47c..6b9b8db 100644
--- a/defs.h
+++ b/defs.h
@@ -4652,7 +4652,11 @@ struct efi_memory_desc_t {
#define STACK_SWITCH_FRAME_REGS 48
#define STACK_FRAME_OVERHEAD 112
-#define EXCP_FRAME_MARKER 0x7265677368657265
+#define EXCP_FRAME_MARKER \
+ (THIS_KERNEL_VERSION >= LINUX(6,1,0) ? \
+ (__BYTE_ORDER == __BIG_ENDIAN ? 0x52454753 : 0x53474552) : \
+ 0x7265677368657265) /* See arch/powerpc/include/asm/ptrace.h: */
+ /* STACK_FRAME_REGS_MARKER */
#define _SECTION_SIZE_BITS 24
#define _MAX_PHYSMEM_BITS 44
diff --git a/ppc64.c b/ppc64.c
index 87f9c4c..f607060 100644
--- a/ppc64.c
+++ b/ppc64.c
@@ -2204,19 +2204,34 @@ ppc64_back_trace(struct gnu_request *req, struct bt_info *bt)
eframe_found = TRUE;
else if (STREQ(req->name, ".ret_from_except"))
eframe_found = TRUE;
- } else if ((newsp - req->sp - STACK_FRAME_OVERHEAD) >=
- sizeof(struct ppc64_pt_regs)) {
- readmem(req->sp+0x60, KVADDR, &marker,
- sizeof(ulong), "stack frame", FAULT_ON_ERROR);
- if (marker == EXCP_FRAME_MARKER)
- eframe_found = TRUE;
+ } else if (THIS_KERNEL_VERSION >= LINUX(6,2,0) && is_ppc64_elf_abi_v2()) {
+ if ((newsp - req->sp - STACK_SWITCH_FRAME_REGS) >=
+ sizeof(struct ppc64_pt_regs)) {
+ readmem(req->sp+0x20, KVADDR, &marker,
+ sizeof(ulong), "stack frame",
+ FAULT_ON_ERROR);
+ }
+ } else {
+ if ((newsp - req->sp - STACK_FRAME_OVERHEAD) >=
+ sizeof(struct ppc64_pt_regs)) {
+ readmem(req->sp+0x60, KVADDR, &marker,
+ sizeof(ulong), "stack frame",
+ FAULT_ON_ERROR);
+ }
}
- if (eframe_found) {
+
+ if (eframe_found || marker == EXCP_FRAME_MARKER) {
char *efrm_str = NULL;
struct ppc64_pt_regs regs;
- readmem(req->sp+STACK_FRAME_OVERHEAD, KVADDR, ®s,
- sizeof(struct ppc64_pt_regs),
- "exception frame", FAULT_ON_ERROR);
+ if (THIS_KERNEL_VERSION >= LINUX(6,2,0) && is_ppc64_elf_abi_v2()) {
+ readmem(req->sp+STACK_SWITCH_FRAME_REGS, KVADDR, ®s,
+ sizeof(struct ppc64_pt_regs),
+ "exception frame", FAULT_ON_ERROR);
+ } else {
+ readmem(req->sp+STACK_FRAME_OVERHEAD, KVADDR, ®s,
+ sizeof(struct ppc64_pt_regs),
+ "exception frame", FAULT_ON_ERROR);
+ }
efrm_str = ppc64_check_eframe(®s);
if (efrm_str) {
@@ -2427,6 +2442,8 @@ ppc64_check_eframe(struct ppc64_pt_regs *regs)
return("Denormalisation");
case 0x1700:
return("Altivec Assist");
+ case 0x3000:
+ return("System Call Vectored");
}
/* No exception frame exists */
--
2.46.2
50
days inactive
50
days old
devel@lists.crash-utility.osci.io
2 comments
1 participants
participants (1)
-
Tao Liu