6:52 a.m.
This series fixes several LoongArch64 crash utility issues seen when
reading CPU register state and unwinding active tasks from vmcore files.
The first two patches fix register extraction from NT_PRSTATUS/crash
notes and initialize the pt_regs.regs offset used for active task register
fallback. The remaining patches improve backtracing across LoongArch
exception frames, fix the stack top boundary check for pt_regs, and avoid
a null eframe_search callback when bt -e is requested.
The NT_PRSTATUS register layout follows the LoongArch kernel
elf_gregset_t/struct pt_regs order: 32 GPRs followed by orig_a0, csr_era,
csr_badvaddr, csr_crmd, csr_prmd, csr_euen, csr_ecfg and csr_estat.
The exception unwinding changes only treat entry symbols that leave a
saved pt_regs at sp as exception frames.
Tested:
- LoongArch 3C6000 platform with vmcore
- make -C crash -j$(nproc) on x86_64 host
- gcc -DLOONGARCH64 -c -o /tmp/loongarch64-final.o loongarch64.c
- gcc -c -o /tmp/loongarch64-final-generic.o loongarch64.c
- git diff --check origin/master..HEAD
Ming Wang (5):
LoongArch64: Fix CPU registers reading from dump notes
LoongArch64: Fix pt_regs initialization for active tasks
LoongArch64: Support backtracing across exception boundaries
LoongArch64: Fix stack frame loop bounds for exception frames
LoongArch64: Add dummy eframe_search to avoid bt -e segfault
loongarch64.c | 78 +++++++++++++++++++++++++++++++++++++++------------
1 file changed, 60 insertions(+), 18 deletions(-)
--
2.43.0
6:52 a.m.
New subject: [PATCH 1/5] LoongArch64: Fix CPU registers reading from dump notes
LoongArch stores the register set exported in NT_PRSTATUS as an
elf_gregset_t. The special registers are laid out as orig_a0, csr_era,
csr_badvaddr, csr_crmd, csr_prmd, csr_euen, csr_ecfg and csr_estat
after the 32 GPRs.
The old crash-local register layout put csr_crmd at the first
special-register slot and placed epc near the end. That made help -r
and active-task unwinding read nonsensical CSR values from dump notes.
Match the local register structure and indexes to the kernel
pt_regs/elf_gregset_t layout. Also handle modern kernels where
crash_notes may refer to the per-cpu note storage directly, and use
NETDUMP ELF notes for libvirt/QEMU dump files when crash_notes does not
provide usable data.
Signed-off-by: Ming Wang <wangming01(a)loongson.cn>
---
loongarch64.c | 38 +++++++++++++++++++++++---------------
1 file changed, 23 insertions(+), 15 deletions(-)
diff --git a/loongarch64.c b/loongarch64.c
index 6437c4aa4656..5adb7d8038db 100644
--- a/loongarch64.c
+++ b/loongarch64.c
@@ -18,20 +18,23 @@
#include <elf.h>
#include "defs.h"
-/* from arch/loongarch/include/asm/ptrace.h */
+/*
+ * This matches the beginning of both kernel struct pt_regs and the
+ * elf_gregset_t layout exported in NT_PRSTATUS notes.
+ */
struct loongarch64_pt_regs {
/* Saved main processor registers. */
unsigned long regs[32];
/* Saved special registers. */
+ unsigned long orig_a0;
+ unsigned long csr_epc;
+ unsigned long csr_badvaddr;
unsigned long csr_crmd;
unsigned long csr_prmd;
unsigned long csr_euen;
unsigned long csr_ecfg;
unsigned long csr_estat;
- unsigned long csr_epc;
- unsigned long csr_badvaddr;
- unsigned long orig_a0;
};
struct loongarch64_unwind_frame {
@@ -108,13 +111,14 @@ typedef struct { ulong pte; } pte_t;
#define LOONGARCH64_EF_RA 1
#define LOONGARCH64_EF_SP 3
#define LOONGARCH64_EF_FP 22
-#define LOONGARCH64_EF_CSR_EPC 32
-#define LOONGARCH64_EF_CSR_BADVADDR 33
-#define LOONGARCH64_EF_CSR_CRMD 34
-#define LOONGARCH64_EF_CSR_PRMD 35
-#define LOONGARCH64_EF_CSR_EUEN 36
-#define LOONGARCH64_EF_CSR_ECFG 37
-#define LOONGARCH64_EF_CSR_ESTAT 38
+#define LOONGARCH64_EF_ORIG_A0 32
+#define LOONGARCH64_EF_CSR_EPC 33
+#define LOONGARCH64_EF_CSR_BADVADDR 34
+#define LOONGARCH64_EF_CSR_CRMD 35
+#define LOONGARCH64_EF_CSR_PRMD 36
+#define LOONGARCH64_EF_CSR_EUEN 37
+#define LOONGARCH64_EF_CSR_ECFG 38
+#define LOONGARCH64_EF_CSR_ESTAT 39
static struct machine_specific loongarch64_machine_specific = { 0 };
@@ -911,8 +915,12 @@ loongarch64_get_crash_notes(void)
/*
* Add __per_cpu_offset for each cpu to form the pointer to the notes
*/
- for (i = 0; i < kt->cpus; i++)
- notes_ptrs[i] = notes_ptrs[kt->cpus-1] + kt->__per_cpu_offset[i];
+ for (i = 0; i < kt->cpus; i++) {
+ if (IS_KVADDR(notes_ptrs[kt->cpus-1]))
+ notes_ptrs[i] = notes_ptrs[kt->cpus-1] + kt->__per_cpu_offset[i];
+ else
+ notes_ptrs[i] = crash_notes + kt->__per_cpu_offset[i];
+ }
}
buf = GETBUF(SIZE(note_buf));
@@ -1009,7 +1017,7 @@ loongarch64_get_elf_notes(void)
struct machine_specific *ms = machdep->machspec;
int i;
- if (!DISKDUMP_DUMPFILE() && !KDUMP_DUMPFILE())
+ if (!DISKDUMP_DUMPFILE() && !KDUMP_DUMPFILE() && !NETDUMP_DUMPFILE())
return FALSE;
panic_task_regs = calloc(kt->cpus, sizeof(*panic_task_regs));
@@ -1022,7 +1030,7 @@ loongarch64_get_elf_notes(void)
if (DISKDUMP_DUMPFILE())
note = diskdump_get_prstatus_percpu(i);
- else if (KDUMP_DUMPFILE())
+ else if (KDUMP_DUMPFILE() || NETDUMP_DUMPFILE())
note = netdump_get_prstatus_percpu(i);
if (!note) {
--
2.43.0
6:52 a.m.
New subject: [PATCH 2/5] LoongArch64: Fix pt_regs initialization for active tasks
The offset of pt_regs.regs was missing, which prevented crash from
falling back to real-time CPU registers for the active task. Initialize
it correctly during architecture initialization.
Signed-off-by: Ming Wang <wangming01(a)loongson.cn>
---
loongarch64.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/loongarch64.c b/loongarch64.c
index 5adb7d8038db..92481f3235b6 100644
--- a/loongarch64.c
+++ b/loongarch64.c
@@ -761,6 +761,8 @@ loongarch64_stackframe_init(void)
ASSIGN_OFFSET(task_struct_thread_reg03) =
task_struct_thread + thread_reg03_sp;
+ MEMBER_OFFSET_INIT(pt_regs_regs, "pt_regs", "regs");
+ STRUCT_SIZE_INIT(pt_regs, "pt_regs");
ASSIGN_OFFSET(task_struct_thread_reg01) =
task_struct_thread + thread_reg01_ra;
--
2.43.0
6:52 a.m.
New subject: [PATCH 3/5] LoongArch64: Support backtracing across exception boundaries
LoongArch exception and syscall entry code saves pt_regs on the kernel
stack before calling the C handlers. When a backtrace reaches one of
these entry symbols, crash needs to recover the saved pt_regs and
continue unwinding from the interrupted context.
Recognize the LoongArch entry symbols that leave a saved pt_regs at sp,
including syscall, exception, interrupt-vector and TLB page-fault paths
used by current upstream kernels and existing Loongnix variants. This
prevents backtraces from stopping at the first entry frame and allows
crash to print the saved pt_regs context without treating ordinary
handler call frames as exception frames.
Signed-off-by: Ming Wang <wangming01(a)loongson.cn>
---
loongarch64.c | 26 +++++++++++++++++++++++++-
1 file changed, 25 insertions(+), 1 deletion(-)
diff --git a/loongarch64.c b/loongarch64.c
index 92481f3235b6..3ec184c908f8 100644
--- a/loongarch64.c
+++ b/loongarch64.c
@@ -707,7 +707,31 @@ loongarch64_is_exception_entry(struct syment *sym)
return STREQ(sym->name, "ret_from_exception") ||
STREQ(sym->name, "ret_from_irq") ||
STREQ(sym->name, "work_resched") ||
- STREQ(sym->name, "handle_sys");
+ STREQ(sym->name, "handle_sys") ||
+ STREQ(sym->name, "handle_syscall") ||
+ STREQ(sym->name, "handle_ade") ||
+ STREQ(sym->name, "handle_ale") ||
+ STREQ(sym->name, "handle_bce") ||
+ STREQ(sym->name, "handle_bp") ||
+ STREQ(sym->name, "handle_fpe") ||
+ STREQ(sym->name, "handle_fpu") ||
+ STREQ(sym->name, "handle_iasub") ||
+ STREQ(sym->name, "handle_ib") ||
+ STREQ(sym->name, "handle_int") ||
+ STREQ(sym->name, "handle_ipe") ||
+ STREQ(sym->name, "handle_lbt") ||
+ STREQ(sym->name, "handle_lsx") ||
+ STREQ(sym->name, "handle_mcheck") ||
+ STREQ(sym->name, "handle_oac") ||
+ STREQ(sym->name, "handle_parchk") ||
+ STREQ(sym->name, "handle_reserved") ||
+ STREQ(sym->name, "handle_ri") ||
+ STREQ(sym->name, "handle_tlb_protect") ||
+ STREQ(sym->name, "tlb_do_page_fault_0") ||
+ STREQ(sym->name, "tlb_do_page_fault_1") ||
+ STREQ(sym->name, "handle_vint") ||
+ STREQ(sym->name, "handle_watch") ||
+ STREQ(sym->name, "handle_lasx");
}
/*
--
2.43.0
6:52 a.m.
New subject: [PATCH 4/5] LoongArch64: Fix stack frame loop bounds for exception frames
Remove the incorrect -32 padding from stacktop limits in loongarch64.c.
In modern LoongArch64 kernels, pt_regs rests exactly at the top of the
stack, and this -32 padding forces the backtrace loop to terminate
prematurely before reaching the exception frame.
Signed-off-by: Ming Wang <wangming01(a)loongson.cn>
---
loongarch64.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/loongarch64.c b/loongarch64.c
index 3ec184c908f8..9980b35bdff1 100644
--- a/loongarch64.c
+++ b/loongarch64.c
@@ -461,7 +461,7 @@ loongarch64_back_trace_cmd(struct bt_info *bt)
previous.pc = current.ra = regs->regs[LOONGARCH64_EF_RA];
}
- while (current.sp <= bt->stacktop - 32 - SIZE(pt_regs)) {
+ while (current.sp <= bt->stacktop - SIZE(pt_regs)) {
struct syment *symbol = NULL;
ulong offset;
@@ -501,7 +501,7 @@ loongarch64_back_trace_cmd(struct bt_info *bt)
* * ret_from_kernel_thread
*/
if (symbol && !STRNEQ(symbol->name, "ret_from") && !offset
&&
- !current.ra && current.sp < bt->stacktop - 32 - SIZE(pt_regs)) {
+ !current.ra && current.sp < bt->stacktop - SIZE(pt_regs)) {
if (CRASHDEBUG(8))
fprintf(fp, "zero offset at %s, try previous symbol\n",
symbol->name);
--
2.43.0
6:52 a.m.
New subject: [PATCH 5/5] LoongArch64: Add dummy eframe_search to avoid bt -e segfault
The lack of machdep->eframe_search initialization causes the crash tool
to segment fault when executing `bt -e`. Provide a stub function that
returns a fatal error indicating the option is not supported, matching
the behavior of other architectures that lack this feature.
Signed-off-by: Ming Wang <wangming01(a)loongson.cn>
---
loongarch64.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/loongarch64.c b/loongarch64.c
index 9980b35bdff1..b0ebf8b1d99e 100644
--- a/loongarch64.c
+++ b/loongarch64.c
@@ -64,6 +64,7 @@ static void loongarch64_dump_backtrace_entry(struct bt_info *bt,
struct loongarch64_unwind_frame *previous, int level);
static void loongarch64_dump_exception_stack(struct bt_info *bt, char *pt_regs);
static int loongarch64_is_exception_entry(struct syment *sym);
+static int loongarch64_eframe_search(struct bt_info *bt);
static void loongarch64_display_full_frame(struct bt_info *bt,
struct loongarch64_unwind_frame *current,
struct loongarch64_unwind_frame *previous);
@@ -1267,6 +1268,7 @@ loongarch64_init(int when)
machdep->kvtop = loongarch64_kvtop;
machdep->cmd_mach = loongarch64_cmd_mach;
machdep->back_trace = loongarch64_back_trace_cmd;
+ machdep->eframe_search = loongarch64_eframe_search;
machdep->get_stack_frame = loongarch64_get_stack_frame;
machdep->vmalloc_start = loongarch64_vmalloc_start;
machdep->processor_speed = loongarch64_processor_speed;
@@ -1389,6 +1391,12 @@ loongarch64_display_regs_from_elf_notes(int cpu, FILE *ofp)
regs->csr_euen);
}
+static int
+loongarch64_eframe_search(struct bt_info *bt)
+{
+ return error(FATAL, "-e option not supported on this architecture\n");
+}
+
#else /* !LOONGARCH64 */
#include "defs.h"
--
2.43.0
6:32 p.m.
New subject: [PATCH 5/5] LoongArch64: Add dummy eframe_search to avoid bt -e segfault
applied:
https://github.com/liutgnu/crash-dev/commit/523981942d8962b5246efa5c92f52...
On Thu, Jun 4, 2026 at 11:54 PM Ming Wang <wangming01(a)loongson.cn> wrote:
The lack of machdep->eframe_search initialization causes the crash tool
to segment fault when executing `bt -e`. Provide a stub function that
returns a fatal error indicating the option is not supported, matching
the behavior of other architectures that lack this feature.
Signed-off-by: Ming Wang <wangming01(a)loongson.cn>
---
loongarch64.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/loongarch64.c b/loongarch64.c
index 9980b35bdff1..b0ebf8b1d99e 100644
--- a/loongarch64.c
+++ b/loongarch64.c
@@ -64,6 +64,7 @@ static void loongarch64_dump_backtrace_entry(struct bt_info *bt,
struct loongarch64_unwind_frame *previous, int level);
static void loongarch64_dump_exception_stack(struct bt_info *bt, char *pt_regs);
static int loongarch64_is_exception_entry(struct syment *sym);
+static int loongarch64_eframe_search(struct bt_info *bt);
static void loongarch64_display_full_frame(struct bt_info *bt,
struct loongarch64_unwind_frame *current,
struct loongarch64_unwind_frame *previous);
@@ -1267,6 +1268,7 @@ loongarch64_init(int when)
machdep->kvtop = loongarch64_kvtop;
machdep->cmd_mach = loongarch64_cmd_mach;
machdep->back_trace = loongarch64_back_trace_cmd;
+ machdep->eframe_search = loongarch64_eframe_search;
machdep->get_stack_frame = loongarch64_get_stack_frame;
machdep->vmalloc_start = loongarch64_vmalloc_start;
machdep->processor_speed = loongarch64_processor_speed;
@@ -1389,6 +1391,12 @@ loongarch64_display_regs_from_elf_notes(int cpu, FILE *ofp)
regs->csr_euen);
}
+static int
+loongarch64_eframe_search(struct bt_info *bt)
+{
+ return error(FATAL, "-e option not supported on this architecture\n");
+}
+
#else /* !LOONGARCH64 */
#include "defs.h"
--
2.43.0
11:27 p.m.
Hi Ming,
I'm OK with this patchset. so ack.
Thanks,
Tao Liu
On Thu, Jun 4, 2026 at 11:54 PM Ming Wang <wangming01(a)loongson.cn> wrote:
This series fixes several LoongArch64 crash utility issues seen when
reading CPU register state and unwinding active tasks from vmcore files.
The first two patches fix register extraction from NT_PRSTATUS/crash
notes and initialize the pt_regs.regs offset used for active task register
fallback. The remaining patches improve backtracing across LoongArch
exception frames, fix the stack top boundary check for pt_regs, and avoid
a null eframe_search callback when bt -e is requested.
The NT_PRSTATUS register layout follows the LoongArch kernel
elf_gregset_t/struct pt_regs order: 32 GPRs followed by orig_a0, csr_era,
csr_badvaddr, csr_crmd, csr_prmd, csr_euen, csr_ecfg and csr_estat.
The exception unwinding changes only treat entry symbols that leave a
saved pt_regs at sp as exception frames.
Tested:
- LoongArch 3C6000 platform with vmcore
- make -C crash -j$(nproc) on x86_64 host
- gcc -DLOONGARCH64 -c -o /tmp/loongarch64-final.o loongarch64.c
- gcc -c -o /tmp/loongarch64-final-generic.o loongarch64.c
- git diff --check origin/master..HEAD
Ming Wang (5):
LoongArch64: Fix CPU registers reading from dump notes
LoongArch64: Fix pt_regs initialization for active tasks
LoongArch64: Support backtracing across exception boundaries
LoongArch64: Fix stack frame loop bounds for exception frames
LoongArch64: Add dummy eframe_search to avoid bt -e segfault
loongarch64.c | 78 +++++++++++++++++++++++++++++++++++++++------------
1 file changed, 60 insertions(+), 18 deletions(-)
--
2.43.0
8:24 p.m.
New subject: [PATCH 0/5] LoongArch64: Fix register notes and exception unwinding
On 6/11/26 12:27, Tao Liu wrote:
Hi Ming,
I'm OK with this patchset. so ack.
Hi Tao Liu,
Thanks for your review.
Thanks,
Ming
Thanks,
Tao Liu
On Thu, Jun 4, 2026 at 11:54 PM Ming Wang <wangming01(a)loongson.cn> wrote:
>
> This series fixes several LoongArch64 crash utility issues seen when
> reading CPU register state and unwinding active tasks from vmcore files.
>
> The first two patches fix register extraction from NT_PRSTATUS/crash
> notes and initialize the pt_regs.regs offset used for active task register
> fallback. The remaining patches improve backtracing across LoongArch
> exception frames, fix the stack top boundary check for pt_regs, and avoid
> a null eframe_search callback when bt -e is requested.
>
> The NT_PRSTATUS register layout follows the LoongArch kernel
> elf_gregset_t/struct pt_regs order: 32 GPRs followed by orig_a0, csr_era,
> csr_badvaddr, csr_crmd, csr_prmd, csr_euen, csr_ecfg and csr_estat.
> The exception unwinding changes only treat entry symbols that leave a
> saved pt_regs at sp as exception frames.
>
> Tested:
> - LoongArch 3C6000 platform with vmcore
> - make -C crash -j$(nproc) on x86_64 host
> - gcc -DLOONGARCH64 -c -o /tmp/loongarch64-final.o loongarch64.c
> - gcc -c -o /tmp/loongarch64-final-generic.o loongarch64.c
> - git diff --check origin/master..HEAD
>
>
> Ming Wang (5):
> LoongArch64: Fix CPU registers reading from dump notes
> LoongArch64: Fix pt_regs initialization for active tasks
> LoongArch64: Support backtracing across exception boundaries
> LoongArch64: Fix stack frame loop bounds for exception frames
> LoongArch64: Add dummy eframe_search to avoid bt -e segfault
>
> loongarch64.c | 78 +++++++++++++++++++++++++++++++++++++++------------
> 1 file changed, 60 insertions(+), 18 deletions(-)
>
> --
> 2.43.0
>
8
days inactive
17
days old
devel@lists.crash-utility.osci.io
8 comments
2 participants
participants (2)
-
Ming Wang -
Tao Liu