[PATCH 0/4] alpha: perf register sampling and user stack dumps

[PATCH 0/4] alpha: perf register sampling and user stack dumps

[Matt Turner]

This lets perf sample registers and dump the user stack on alpha, which
perf record --call-graph dwarf needs to unwind user space.

r9-r15 are not in pt_regs, so they cannot be sampled. An unwinder
therefore cannot get past a first frame that uses $15 as its frame
pointer (-O0 code or alloca).

The perf tool side is a separate patch for perf-tools-next. libdw also
needs alpha unwinding support, which I am sending to elfutils-devel.

---
Matt Turner (4):
      alpha: fix stale user_stack_pointer()
      alpha: stop storing the user stack pointer in pt_regs
      alpha: use user_stack_pointer() in KSTK_ESP() and dump_elf_thread()
      alpha: add HAVE_PERF_REGS and HAVE_PERF_USER_STACK_DUMP support

 .../features/perf/perf-regs/arch-support.txt       |  2 +-
 .../features/perf/perf-stackdump/arch-support.txt  |  2 +-
 arch/alpha/Kconfig                                 |  2 +
 arch/alpha/include/asm/processor.h                 |  3 +-
 arch/alpha/include/asm/ptrace.h                    |  7 +--
 arch/alpha/include/uapi/asm/perf_regs.h            | 45 +++++++++++++++++
 arch/alpha/include/uapi/asm/ptrace.h               |  2 +-
 arch/alpha/kernel/Makefile                         |  1 +
 arch/alpha/kernel/asm-offsets.c                    |  1 -
 arch/alpha/kernel/entry.S                          |  5 --
 arch/alpha/kernel/perf_regs.c                      | 56 ++++++++++++++++++++++
 arch/alpha/kernel/process.c                        |  2 +-
 arch/alpha/kernel/ptrace.c                         | 12 +++++
 arch/alpha/kernel/traps.c                          |  8 ----
 14 files changed, 122 insertions(+), 26 deletions(-)
---
base-commit: 8f5dbfb93440d1aba216a1a34369a65654b94be9
change-id: 20261001-alpha-perf-regs-v3-7e23508c1ca2

Best regards,
-- 
Matt Turner <mattst88@gmail.com>

[PATCH 1/4] alpha: fix stale user_stack_pointer()

[Matt Turner]

PALcode keeps the user stack pointer in the PCB, not the trap frame, so
pt_regs->usp is only valid where entry code stores it: entSys,
do_entArith() and do_entIF(). Interrupts, page faults and unaligned
traps leave it stale, and ptrace and /proc/<pid>/syscall read it
through user_stack_pointer().

A user-mode pt_regs is at the top of its task's stack, so find the
thread_info from it and use rdusp() for current or pcb.usp otherwise,
as KSTK_ESP() and the regset code already do.

Fixes: 8a2857bc2b30 ("alpha: enable regset-based ptrace and core dumps")
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Matt Turner <mattst88@gmail.com>
---
 arch/alpha/include/asm/ptrace.h |  7 +------
 arch/alpha/kernel/ptrace.c      | 12 ++++++++++++
 2 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/arch/alpha/include/asm/ptrace.h b/arch/alpha/include/asm/ptrace.h
index 430e8dc27ff7..45d1229344b6 100644
--- a/arch/alpha/include/asm/ptrace.h
+++ b/arch/alpha/include/asm/ptrace.h
@@ -26,12 +26,7 @@ static inline unsigned long regs_return_value(struct pt_regs *regs)
 	return regs->r0;
 }
 
-/* Helpers for working with the user stack pointer */
-static inline unsigned long user_stack_pointer(struct pt_regs *regs)
-{
-	/* Valid for user-mode regs */
-	return regs->usp;
-}
+unsigned long user_stack_pointer(struct pt_regs *regs);
 
 static __always_inline bool regs_irqs_disabled(struct pt_regs *regs)
 {
diff --git a/arch/alpha/kernel/ptrace.c b/arch/alpha/kernel/ptrace.c
index e1d8c2f69688..dba7e9905cfa 100644
--- a/arch/alpha/kernel/ptrace.c
+++ b/arch/alpha/kernel/ptrace.c
@@ -440,6 +440,18 @@ ptrace_cancel_bpt(struct task_struct * child)
 	return (nsaved != 0);
 }
 
+/*
+ * PALcode keeps the user stack pointer in the PCB, not the trap frame.
+ * A user-mode pt_regs sits at the top of its task's stack, so the
+ * thread_info is at the base.
+ */
+unsigned long user_stack_pointer(struct pt_regs *regs)
+{
+	struct thread_info *ti = (void *)((unsigned long)regs & -THREAD_SIZE);
+
+	return ti == current_thread_info() ? rdusp() : ti->pcb.usp;
+}
+
 void user_enable_single_step(struct task_struct *child)
 {
 	/* Mark single stepping.  */

-- 
2.55.0

[PATCH 2/4] alpha: stop storing the user stack pointer in pt_regs

[Matt Turner]

Nothing reads pt_regs->usp now that user_stack_pointer() uses the PCB.
Drop the stores, which saves a PAL_rdusp call on every syscall, and
give the field back its old name.

Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Matt Turner <mattst88@gmail.com>
---
 arch/alpha/include/uapi/asm/ptrace.h | 2 +-
 arch/alpha/kernel/asm-offsets.c      | 1 -
 arch/alpha/kernel/entry.S            | 5 -----
 arch/alpha/kernel/traps.c            | 8 --------
 4 files changed, 1 insertion(+), 15 deletions(-)

diff --git a/arch/alpha/include/uapi/asm/ptrace.h b/arch/alpha/include/uapi/asm/ptrace.h
index 9d86b2a1526e..72ed913a910f 100644
--- a/arch/alpha/include/uapi/asm/ptrace.h
+++ b/arch/alpha/include/uapi/asm/ptrace.h
@@ -43,7 +43,7 @@ struct pt_regs {
 	unsigned long trap_a1;
 	unsigned long trap_a2;
 /* This makes the stack 16-byte aligned as GCC expects */
-	unsigned long usp;
+	unsigned long __pad0;
 /* These are saved by PAL-code: */
 	unsigned long ps;
 	unsigned long pc;
diff --git a/arch/alpha/kernel/asm-offsets.c b/arch/alpha/kernel/asm-offsets.c
index c89792c86044..22ed3c975b1c 100644
--- a/arch/alpha/kernel/asm-offsets.c
+++ b/arch/alpha/kernel/asm-offsets.c
@@ -30,7 +30,6 @@ static void __used foo(void)
 
 	DEFINE(HAE_CACHE, offsetof(struct alpha_machine_vector, hae_cache));
 	DEFINE(HAE_REG, offsetof(struct alpha_machine_vector, hae_register));
-	DEFINE(PT_REGS_USP, offsetof(struct pt_regs, usp));
 
 	DEFINE(TI_FLAGS,        offsetof(struct thread_info, flags));
 	DEFINE(TI_SYSCALL_WORK, offsetof(struct thread_info, syscall_work));
diff --git a/arch/alpha/kernel/entry.S b/arch/alpha/kernel/entry.S
index bfa116d455fb..790389a15c3c 100644
--- a/arch/alpha/kernel/entry.S
+++ b/arch/alpha/kernel/entry.S
@@ -476,11 +476,6 @@ entSys:
 	ldq	$1, 0($sp)	/* syscall nr from saved r0 */
 	stq	$1, 8($sp)	/* regs->r1 = shadow syscall nr */
 
-	mov	$0, $8
-	call_pal PAL_rdusp
-	stq	$0, PT_REGS_USP($sp)
-	mov	$8, $0
-
 	lda	$8, 0x3fff
 	bic	$sp, $8, $8
 
diff --git a/arch/alpha/kernel/traps.c b/arch/alpha/kernel/traps.c
index cfaef96c4ce8..5882c65b7ef1 100644
--- a/arch/alpha/kernel/traps.c
+++ b/arch/alpha/kernel/traps.c
@@ -30,12 +30,6 @@
 
 #include "proto.h"
 
-static __always_inline void alpha_snapshot_usp(struct pt_regs *regs)
-{
-	if (user_mode(regs))
-		regs->usp = rdusp();
-}
-
 void
 dik_show_regs(struct pt_regs *regs, unsigned long *r9_15)
 {
@@ -186,7 +180,6 @@ do_entArith(unsigned long summary, unsigned long write_mask,
 {
 	long si_code = FPE_FLTINV;
 
-	alpha_snapshot_usp(regs);
 	if (summary & 1) {
 		/* Software-completion summary bit is set, so try to
 		   emulate the instruction.  If the processor supports
@@ -208,7 +201,6 @@ do_entIF(unsigned long type, struct pt_regs *regs)
 {
 	int signo, code;
 
-	alpha_snapshot_usp(regs);
 	if (type == 3) { /* FEN fault */
 		/* Irritating users can call PAL_clrfen to disable the
 		   FPU for the process.  The kernel will then trap in

-- 
2.55.0

[PATCH 3/4] alpha: use user_stack_pointer() in KSTK_ESP() and dump_elf_thread()

[Matt Turner]

Both open-code the same PCB lookup.

Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Matt Turner <mattst88@gmail.com>
---
 arch/alpha/include/asm/processor.h | 3 +--
 arch/alpha/kernel/process.c        | 2 +-
 2 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/arch/alpha/include/asm/processor.h b/arch/alpha/include/asm/processor.h
index c461a26d47b3..73251cd671e0 100644
--- a/arch/alpha/include/asm/processor.h
+++ b/arch/alpha/include/asm/processor.h
@@ -43,8 +43,7 @@ unsigned long __get_wchan(struct task_struct *p);
 
 #define KSTK_EIP(tsk) (task_pt_regs(tsk)->pc)
 
-#define KSTK_ESP(tsk) \
-  ((tsk) == current ? rdusp() : task_thread_info(tsk)->pcb.usp)
+#define KSTK_ESP(tsk) user_stack_pointer(task_pt_regs(tsk))
 
 #define cpu_relax()	barrier()
 
diff --git a/arch/alpha/kernel/process.c b/arch/alpha/kernel/process.c
index 06522451f018..7861c9e767b7 100644
--- a/arch/alpha/kernel/process.c
+++ b/arch/alpha/kernel/process.c
@@ -318,7 +318,7 @@ dump_elf_thread(elf_greg_t *dest, struct pt_regs *pt, struct thread_info *ti)
 	dest[27] = pt->r27;
 	dest[28] = pt->r28;
 	dest[29] = pt->gp;
-	dest[30] = ti == current_thread_info() ? rdusp() : ti->pcb.usp;
+	dest[30] = user_stack_pointer(pt);
 	dest[31] = pt->pc;
 
 	/* Once upon a time this was the PS value.  Which is stupid

-- 
2.55.0

[PATCH 4/4] alpha: add HAVE_PERF_REGS and HAVE_PERF_USER_STACK_DUMP support

[Matt Turner]

Let perf sample registers and dump the user stack, which is what the
perf tool needs for --call-graph dwarf.

The register indices follow the hardware numbering, so r0-r30 map
directly to DWARF register numbers, with pc after them. r9-r15
are callee-saved and live in switch_stack rather than pt_regs, so they
cannot be sampled. Without r15 an unwinder cannot recover the CFA of a
first frame that uses $15 as its frame pointer.

sp is not in pt_regs either: for user frames it comes from
user_stack_pointer(), and for kernel frames (PERF_SAMPLE_REGS_INTR) it
is the end of the trap frame, as in stacktrace.c.

Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Matt Turner <mattst88@gmail.com>
---
 .../features/perf/perf-regs/arch-support.txt       |  2 +-
 .../features/perf/perf-stackdump/arch-support.txt  |  2 +-
 arch/alpha/Kconfig                                 |  2 +
 arch/alpha/include/uapi/asm/perf_regs.h            | 45 +++++++++++++++++
 arch/alpha/kernel/Makefile                         |  1 +
 arch/alpha/kernel/perf_regs.c                      | 56 ++++++++++++++++++++++
 6 files changed, 106 insertions(+), 2 deletions(-)

diff --git a/Documentation/features/perf/perf-regs/arch-support.txt b/Documentation/features/perf/perf-regs/arch-support.txt
index 4c9a5a012075..7d736d678fd4 100644
--- a/Documentation/features/perf/perf-regs/arch-support.txt
+++ b/Documentation/features/perf/perf-regs/arch-support.txt
@@ -6,7 +6,7 @@
     -----------------------
     |         arch |status|
     -----------------------
-    |       alpha: | TODO |
+    |       alpha: |  ok  |
     |         arc: | TODO |
     |         arm: |  ok  |
     |       arm64: |  ok  |
diff --git a/Documentation/features/perf/perf-stackdump/arch-support.txt b/Documentation/features/perf/perf-stackdump/arch-support.txt
index dd362b5cb638..17b84ea7769d 100644
--- a/Documentation/features/perf/perf-stackdump/arch-support.txt
+++ b/Documentation/features/perf/perf-stackdump/arch-support.txt
@@ -6,7 +6,7 @@
     -----------------------
     |         arch |status|
     -----------------------
-    |       alpha: | TODO |
+    |       alpha: |  ok  |
     |         arc: | TODO |
     |         arm: |  ok  |
     |       arm64: |  ok  |
diff --git a/arch/alpha/Kconfig b/arch/alpha/Kconfig
index b2c359b947f6..8fafbfabd1e5 100644
--- a/arch/alpha/Kconfig
+++ b/arch/alpha/Kconfig
@@ -41,6 +41,8 @@ config ALPHA
 	select HAVE_PAGE_SIZE_8KB
 	select HAVE_PCSPKR_PLATFORM
 	select HAVE_PERF_EVENTS
+	select HAVE_PERF_REGS
+	select HAVE_PERF_USER_STACK_DUMP
 	select NEED_DMA_MAP_STATE
 	select NEED_SG_DMA_LENGTH
 	select GENERIC_IRQ_PROBE
diff --git a/arch/alpha/include/uapi/asm/perf_regs.h b/arch/alpha/include/uapi/asm/perf_regs.h
new file mode 100644
index 000000000000..da96a31b12fb
--- /dev/null
+++ b/arch/alpha/include/uapi/asm/perf_regs.h
@@ -0,0 +1,45 @@
+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
+#ifndef _UAPI_ASM_ALPHA_PERF_REGS_H
+#define _UAPI_ASM_ALPHA_PERF_REGS_H
+
+/*
+ * Indices 0-30 match the hardware (and DWARF) register numbers.
+ * r9-r15 are not saved in pt_regs and cannot be sampled.
+ */
+enum perf_event_alpha_regs {
+	PERF_REG_ALPHA_R0,
+	PERF_REG_ALPHA_R1,
+	PERF_REG_ALPHA_R2,
+	PERF_REG_ALPHA_R3,
+	PERF_REG_ALPHA_R4,
+	PERF_REG_ALPHA_R5,
+	PERF_REG_ALPHA_R6,
+	PERF_REG_ALPHA_R7,
+	PERF_REG_ALPHA_R8,
+	PERF_REG_ALPHA_R9,
+	PERF_REG_ALPHA_R10,
+	PERF_REG_ALPHA_R11,
+	PERF_REG_ALPHA_R12,
+	PERF_REG_ALPHA_R13,
+	PERF_REG_ALPHA_R14,
+	PERF_REG_ALPHA_R15,
+	PERF_REG_ALPHA_R16,
+	PERF_REG_ALPHA_R17,
+	PERF_REG_ALPHA_R18,
+	PERF_REG_ALPHA_R19,
+	PERF_REG_ALPHA_R20,
+	PERF_REG_ALPHA_R21,
+	PERF_REG_ALPHA_R22,
+	PERF_REG_ALPHA_R23,
+	PERF_REG_ALPHA_R24,
+	PERF_REG_ALPHA_R25,
+	PERF_REG_ALPHA_R26,
+	PERF_REG_ALPHA_R27,
+	PERF_REG_ALPHA_R28,
+	PERF_REG_ALPHA_GP,
+	PERF_REG_ALPHA_SP,
+	PERF_REG_ALPHA_PC,
+	PERF_REG_ALPHA_MAX,
+};
+
+#endif /* _UAPI_ASM_ALPHA_PERF_REGS_H */
diff --git a/arch/alpha/kernel/Makefile b/arch/alpha/kernel/Makefile
index 4ea5c189e60e..0c04b89676ed 100644
--- a/arch/alpha/kernel/Makefile
+++ b/arch/alpha/kernel/Makefile
@@ -18,6 +18,7 @@ obj-$(CONFIG_PCI)	+= pci.o pci_iommu.o pci-sysfs.o
 obj-$(CONFIG_SRM_ENV)	+= srm_env.o
 obj-$(CONFIG_MODULES)	+= module.o
 obj-$(CONFIG_PERF_EVENTS) += perf_event.o
+obj-$(CONFIG_HAVE_PERF_REGS) += perf_regs.o
 obj-$(CONFIG_RTC_DRV_ALPHA) += rtc.o
 obj-$(CONFIG_AUDIT)	+= audit.o
 
diff --git a/arch/alpha/kernel/perf_regs.c b/arch/alpha/kernel/perf_regs.c
new file mode 100644
index 000000000000..10afe4e205e9
--- /dev/null
+++ b/arch/alpha/kernel/perf_regs.c
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/bits.h>
+#include <linux/perf_event.h>
+#include <asm/perf_regs.h>
+
+#define REG(idx, field) \
+	[PERF_REG_ALPHA_##idx] = offsetof(struct pt_regs, field)
+
+#define REG_RESERVED	(~GENMASK_ULL(PERF_REG_ALPHA_MAX - 1, 0) | \
+			 GENMASK_ULL(PERF_REG_ALPHA_R15, PERF_REG_ALPHA_R9))
+
+static const unsigned int pt_regs_offset[PERF_REG_ALPHA_MAX] = {
+	REG(R0, r0),	REG(R1, r1),	REG(R2, r2),	REG(R3, r3),
+	REG(R4, r4),	REG(R5, r5),	REG(R6, r6),	REG(R7, r7),
+	REG(R8, r8),	REG(R16, r16),	REG(R17, r17),	REG(R18, r18),
+	REG(R19, r19),	REG(R20, r20),	REG(R21, r21),	REG(R22, r22),
+	REG(R23, r23),	REG(R24, r24),	REG(R25, r25),	REG(R26, r26),
+	REG(R27, r27),	REG(R28, r28),	REG(GP, gp),	REG(PC, pc),
+};
+
+u64 perf_reg_value(struct pt_regs *regs, int idx)
+{
+	if (idx == PERF_REG_ALPHA_SP) {
+		if (user_mode(regs))
+			return user_stack_pointer(regs);
+		/* Same kernel sp approximation as stacktrace.c. */
+		return (unsigned long)(regs + 1);
+	}
+
+	if (WARN_ON_ONCE((u32)idx >= PERF_REG_ALPHA_MAX ||
+			 BIT_ULL(idx) & REG_RESERVED))
+		return 0;
+
+	return *(unsigned long *)((void *)regs + pt_regs_offset[idx]);
+}
+
+int perf_reg_validate(u64 mask)
+{
+	if (!mask || mask & REG_RESERVED)
+		return -EINVAL;
+
+	return 0;
+}
+
+u64 perf_reg_abi(struct task_struct *task)
+{
+	return PERF_SAMPLE_REGS_ABI_64;
+}
+
+void perf_get_regs_user(struct perf_regs *regs_user,
+			struct pt_regs *regs)
+{
+	regs_user->regs = task_pt_regs(current);
+	regs_user->abi = perf_reg_abi(current);
+}

-- 
2.55.0

Re: [PATCH 0/4] alpha: perf register sampling and user stack dumps

[Ian Rogers]

On Sun, Oct 11, 2026 at 1:16 PM Matt Turner <mattst88@gmail.com> wrote:
>
> This lets perf sample registers and dump the user stack on alpha, which
> perf record --call-graph dwarf needs to unwind user space.
>
> r9-r15 are not in pt_regs, so they cannot be sampled. An unwinder
> therefore cannot get past a first frame that uses $15 as its frame
> pointer (-O0 code or alloca).
>
> The perf tool side is a separate patch for perf-tools-next. libdw also
> needs alpha unwinding support, which I am sending to elfutils-devel.
>
> ---
> Matt Turner (4):
>       alpha: fix stale user_stack_pointer()
>       alpha: stop storing the user stack pointer in pt_regs
>       alpha: use user_stack_pointer() in KSTK_ESP() and dump_elf_thread()
>       alpha: add HAVE_PERF_REGS and HAVE_PERF_USER_STACK_DUMP support

Acked-by: Ian Rogers <irogers@google.com>

Thanks!
Ian

>  .../features/perf/perf-regs/arch-support.txt       |  2 +-
>  .../features/perf/perf-stackdump/arch-support.txt  |  2 +-
>  arch/alpha/Kconfig                                 |  2 +
>  arch/alpha/include/asm/processor.h                 |  3 +-
>  arch/alpha/include/asm/ptrace.h                    |  7 +--
>  arch/alpha/include/uapi/asm/perf_regs.h            | 45 +++++++++++++++++
>  arch/alpha/include/uapi/asm/ptrace.h               |  2 +-
>  arch/alpha/kernel/Makefile                         |  1 +
>  arch/alpha/kernel/asm-offsets.c                    |  1 -
>  arch/alpha/kernel/entry.S                          |  5 --
>  arch/alpha/kernel/perf_regs.c                      | 56 ++++++++++++++++++++++
>  arch/alpha/kernel/process.c                        |  2 +-
>  arch/alpha/kernel/ptrace.c                         | 12 +++++
>  arch/alpha/kernel/traps.c                          |  8 ----
>  14 files changed, 122 insertions(+), 26 deletions(-)
> ---
> base-commit: 8f5dbfb93440d1aba216a1a34369a65654b94be9
> change-id: 20261001-alpha-perf-regs-v3-7e23508c1ca2
>
> Best regards,
> --
> Matt Turner <mattst88@gmail.com>
>
>

Re: [PATCH 0/4] alpha: perf register sampling and user stack dumps

[Magnus Lindholm]

Hi Matt,

On Sun, Oct 11, 2026 at 10:16 PM Matt Turner <mattst88@gmail.com> wrote:
>
> This lets perf sample registers and dump the user stack on alpha, which
> perf record --call-graph dwarf needs to unwind user space.
>
> r9-r15 are not in pt_regs, so they cannot be sampled. An unwinder
> therefore cannot get past a first frame that uses $15 as its frame
> pointer (-O0 code or alloca).
>
> The perf tool side is a separate patch for perf-tools-next. libdw also
> needs alpha unwinding support, which I am sending to elfutils-devel.
>
> ---
> Matt Turner (4):
>       alpha: fix stale user_stack_pointer()
>       alpha: stop storing the user stack pointer in pt_regs
>       alpha: use user_stack_pointer() in KSTK_ESP() and dump_elf_thread()
>       alpha: add HAVE_PERF_REGS and HAVE_PERF_USER_STACK_DUMP support
>
>  .../features/perf/perf-regs/arch-support.txt       |  2 +-
>  .../features/perf/perf-stackdump/arch-support.txt  |  2 +-
>  arch/alpha/Kconfig                                 |  2 +
>  arch/alpha/include/asm/processor.h                 |  3 +-
>  arch/alpha/include/asm/ptrace.h                    |  7 +--
>  arch/alpha/include/uapi/asm/perf_regs.h            | 45 +++++++++++++++++
>  arch/alpha/include/uapi/asm/ptrace.h               |  2 +-
>  arch/alpha/kernel/Makefile                         |  1 +
>  arch/alpha/kernel/asm-offsets.c                    |  1 -
>  arch/alpha/kernel/entry.S                          |  5 --
>  arch/alpha/kernel/perf_regs.c                      | 56 ++++++++++++++++++++++
>  arch/alpha/kernel/process.c                        |  2 +-
>  arch/alpha/kernel/ptrace.c                         | 12 +++++
>  arch/alpha/kernel/traps.c                          |  8 ----
>  14 files changed, 122 insertions(+), 26 deletions(-)
> ---
> base-commit: 8f5dbfb93440d1aba216a1a34369a65654b94be9
> change-id: 20261001-alpha-perf-regs-v3-7e23508c1ca2
>
> Best regards,
> --
> Matt Turner <mattst88@gmail.com>
>

The four kernel patches apply cleanly on my for-next branch. I built and
booted the series on my ES40 (EV67/Tsunami), and tested software-clock,
page-fault and hardware-cycle sampling, register masks and user stack
dumps across all three online CPUs. The targeted tests passed.

I have a few review comments(I'll respond to the individual patches).
The separate perf patch records and decodes the samples, but
I cannot yet confirm DWARF unwinding: the installed elfutils 0.195 lacks
the required Alpha support. Could you point me to that patch as well?

Thanks,
Magnus

Re: [PATCH 1/4] alpha: fix stale user_stack_pointer()

[Magnus Lindholm]

Hi Matt,

On Sun, Oct 11, 2026 at 10:16 PM Matt Turner <mattst88@gmail.com> wrote:
>
> PALcode keeps the user stack pointer in the PCB, not the trap frame, so
> pt_regs->usp is only valid where entry code stores it: entSys,
> do_entArith() and do_entIF(). Interrupts, page faults and unaligned
> traps leave it stale, and ptrace and /proc/<pid>/syscall read it
> through user_stack_pointer().
>
> A user-mode pt_regs is at the top of its task's stack, so find the
> thread_info from it and use rdusp() for current or pcb.usp otherwise,
> as KSTK_ESP() and the regset code already do.
>
> Fixes: 8a2857bc2b30 ("alpha: enable regset-based ptrace and core dumps")
> Assisted-by: Claude:claude-opus-5-5
> Signed-off-by: Matt Turner <mattst88@gmail.com>
> ---
>  arch/alpha/include/asm/ptrace.h |  7 +------
>  arch/alpha/kernel/ptrace.c      | 12 ++++++++++++
>  2 files changed, 13 insertions(+), 6 deletions(-)
>
> diff --git a/arch/alpha/include/asm/ptrace.h b/arch/alpha/include/asm/ptrace.h
> index 430e8dc27ff7..45d1229344b6 100644
> --- a/arch/alpha/include/asm/ptrace.h
> +++ b/arch/alpha/include/asm/ptrace.h
> @@ -26,12 +26,7 @@ static inline unsigned long regs_return_value(struct pt_regs *regs)
>         return regs->r0;
>  }
>
> -/* Helpers for working with the user stack pointer */
> -static inline unsigned long user_stack_pointer(struct pt_regs *regs)
> -{
> -       /* Valid for user-mode regs */
> -       return regs->usp;
> -}
> +unsigned long user_stack_pointer(struct pt_regs *regs);
>
>  static __always_inline bool regs_irqs_disabled(struct pt_regs *regs)
>  {
> diff --git a/arch/alpha/kernel/ptrace.c b/arch/alpha/kernel/ptrace.c
> index e1d8c2f69688..dba7e9905cfa 100644
> --- a/arch/alpha/kernel/ptrace.c
> +++ b/arch/alpha/kernel/ptrace.c
> @@ -440,6 +440,18 @@ ptrace_cancel_bpt(struct task_struct * child)
>         return (nsaved != 0);
>  }
>
> +/*
> + * PALcode keeps the user stack pointer in the PCB, not the trap frame.
> + * A user-mode pt_regs sits at the top of its task's stack, so the
> + * thread_info is at the base.
> + */
> +unsigned long user_stack_pointer(struct pt_regs *regs)
> +{
> +       struct thread_info *ti = (void *)((unsigned long)regs & -THREAD_SIZE);
> +
> +       return ti == current_thread_info() ? rdusp() : ti->pcb.usp;
> +}
> +
>  void user_enable_single_step(struct task_struct *child)
>  {
>         /* Mark single stepping.  */
>
> --
> 2.55.0
>

Nice work!

Reviewed-by: Magnus Lindholm <linmag7@gmail.com>
Tested-by: Magnus Lindholm <linmag7@gmail.com>

Re: [PATCH 2/4] alpha: stop storing the user stack pointer in pt_regs

[Magnus Lindholm]

Hi Matt,


On Sun, Oct 11, 2026 at 10:16 PM Matt Turner <mattst88@gmail.com> wrote:
>
> Nothing reads pt_regs->usp now that user_stack_pointer() uses the PCB.
> Drop the stores, which saves a PAL_rdusp call on every syscall, and
> give the field back its old name.
>
> Assisted-by: Claude:claude-opus-5-5
> Signed-off-by: Matt Turner <mattst88@gmail.com>
> ---
>  arch/alpha/include/uapi/asm/ptrace.h | 2 +-
>  arch/alpha/kernel/asm-offsets.c      | 1 -
>  arch/alpha/kernel/entry.S            | 5 -----
>  arch/alpha/kernel/traps.c            | 8 --------
>  4 files changed, 1 insertion(+), 15 deletions(-)
>
> diff --git a/arch/alpha/include/uapi/asm/ptrace.h b/arch/alpha/include/uapi/asm/ptrace.h
> index 9d86b2a1526e..72ed913a910f 100644
> --- a/arch/alpha/include/uapi/asm/ptrace.h
> +++ b/arch/alpha/include/uapi/asm/ptrace.h
> @@ -43,7 +43,7 @@ struct pt_regs {
>         unsigned long trap_a1;
>         unsigned long trap_a2;
>  /* This makes the stack 16-byte aligned as GCC expects */
> -       unsigned long usp;
> +       unsigned long __pad0;
>  /* These are saved by PAL-code: */
>         unsigned long ps;
>         unsigned long pc;
> diff --git a/arch/alpha/kernel/asm-offsets.c b/arch/alpha/kernel/asm-offsets.c
> index c89792c86044..22ed3c975b1c 100644
> --- a/arch/alpha/kernel/asm-offsets.c
> +++ b/arch/alpha/kernel/asm-offsets.c
> @@ -30,7 +30,6 @@ static void __used foo(void)
>
>         DEFINE(HAE_CACHE, offsetof(struct alpha_machine_vector, hae_cache));
>         DEFINE(HAE_REG, offsetof(struct alpha_machine_vector, hae_register));
> -       DEFINE(PT_REGS_USP, offsetof(struct pt_regs, usp));
>
>         DEFINE(TI_FLAGS,        offsetof(struct thread_info, flags));
>         DEFINE(TI_SYSCALL_WORK, offsetof(struct thread_info, syscall_work));
> diff --git a/arch/alpha/kernel/entry.S b/arch/alpha/kernel/entry.S
> index bfa116d455fb..790389a15c3c 100644
> --- a/arch/alpha/kernel/entry.S
> +++ b/arch/alpha/kernel/entry.S
> @@ -476,11 +476,6 @@ entSys:
>         ldq     $1, 0($sp)      /* syscall nr from saved r0 */
>         stq     $1, 8($sp)      /* regs->r1 = shadow syscall nr */
>
> -       mov     $0, $8
> -       call_pal PAL_rdusp
> -       stq     $0, PT_REGS_USP($sp)
> -       mov     $8, $0
> -
>         lda     $8, 0x3fff
>         bic     $sp, $8, $8
>
> diff --git a/arch/alpha/kernel/traps.c b/arch/alpha/kernel/traps.c
> index cfaef96c4ce8..5882c65b7ef1 100644
> --- a/arch/alpha/kernel/traps.c
> +++ b/arch/alpha/kernel/traps.c
> @@ -30,12 +30,6 @@
>
>  #include "proto.h"
>
> -static __always_inline void alpha_snapshot_usp(struct pt_regs *regs)
> -{
> -       if (user_mode(regs))
> -               regs->usp = rdusp();
> -}
> -
>  void
>  dik_show_regs(struct pt_regs *regs, unsigned long *r9_15)
>  {
> @@ -186,7 +180,6 @@ do_entArith(unsigned long summary, unsigned long write_mask,
>  {
>         long si_code = FPE_FLTINV;
>
> -       alpha_snapshot_usp(regs);
>         if (summary & 1) {
>                 /* Software-completion summary bit is set, so try to
>                    emulate the instruction.  If the processor supports
> @@ -208,7 +201,6 @@ do_entIF(unsigned long type, struct pt_regs *regs)
>  {
>         int signo, code;
>
> -       alpha_snapshot_usp(regs);
>         if (type == 3) { /* FEN fault */
>                 /* Irritating users can call PAL_clrfen to disable the
>                    FPU for the process.  The kernel will then trap in
>
> --
> 2.55.0
>

Looks good to me, I see the __pad0 is back again in pt_regs
brings back memories from tracking down the stack alignment
issues a while back.

Reviewed-by: Magnus Lindholm <linmag7@gmail.com>
Tested-by: Magnus Lindholm <linmag7@gmail.com>

Re: [PATCH 3/4] alpha: use user_stack_pointer() in KSTK_ESP() and dump_elf_thread()

[Magnus Lindholm]

Hi Matt,

On Sun, Oct 11, 2026 at 10:16 PM Matt Turner <mattst88@gmail.com> wrote:
>
> Both open-code the same PCB lookup.
>
> Assisted-by: Claude:claude-opus-5-5
> Signed-off-by: Matt Turner <mattst88@gmail.com>
> ---
>  arch/alpha/include/asm/processor.h | 3 +--
>  arch/alpha/kernel/process.c        | 2 +-
>  2 files changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/arch/alpha/include/asm/processor.h b/arch/alpha/include/asm/processor.h
> index c461a26d47b3..73251cd671e0 100644
> --- a/arch/alpha/include/asm/processor.h
> +++ b/arch/alpha/include/asm/processor.h
> @@ -43,8 +43,7 @@ unsigned long __get_wchan(struct task_struct *p);
>
>  #define KSTK_EIP(tsk) (task_pt_regs(tsk)->pc)
>
> -#define KSTK_ESP(tsk) \
> -  ((tsk) == current ? rdusp() : task_thread_info(tsk)->pcb.usp)
> +#define KSTK_ESP(tsk) user_stack_pointer(task_pt_regs(tsk))
>
>  #define cpu_relax()    barrier()
>
> diff --git a/arch/alpha/kernel/process.c b/arch/alpha/kernel/process.c
> index 06522451f018..7861c9e767b7 100644
> --- a/arch/alpha/kernel/process.c
> +++ b/arch/alpha/kernel/process.c
> @@ -318,7 +318,7 @@ dump_elf_thread(elf_greg_t *dest, struct pt_regs *pt, struct thread_info *ti)
>         dest[27] = pt->r27;
>         dest[28] = pt->r28;
>         dest[29] = pt->gp;
> -       dest[30] = ti == current_thread_info() ? rdusp() : ti->pcb.usp;
> +       dest[30] = user_stack_pointer(pt);
>         dest[31] = pt->pc;
>
>         /* Once upon a time this was the PS value.  Which is stupid
>
> --
> 2.55.0
>

Looks good to me.

Tested-by: Magnus Lindholm <linmag7@gmail.com>
Reviewed-by: Magnus Lindholm <linmag7@gmail.com>

Re: [PATCH 4/4] alpha: add HAVE_PERF_REGS and HAVE_PERF_USER_STACK_DUMP support

[Magnus Lindholm]

Hi Matt,

On Sun, Oct 11, 2026 at 10:16 PM Matt Turner <mattst88@gmail.com> wrote:
>
> Let perf sample registers and dump the user stack, which is what the
> perf tool needs for --call-graph dwarf.
>
> The register indices follow the hardware numbering, so r0-r30 map
> directly to DWARF register numbers, with pc after them. r9-r15
> are callee-saved and live in switch_stack rather than pt_regs, so they
> cannot be sampled. Without r15 an unwinder cannot recover the CFA of a
> first frame that uses $15 as its frame pointer.
>
> sp is not in pt_regs either: for user frames it comes from
> user_stack_pointer(), and for kernel frames (PERF_SAMPLE_REGS_INTR) it
> is the end of the trap frame, as in stacktrace.c.
>
> Assisted-by: Claude:claude-opus-5-5
> Signed-off-by: Matt Turner <mattst88@gmail.com>
> ---
>  .../features/perf/perf-regs/arch-support.txt       |  2 +-
>  .../features/perf/perf-stackdump/arch-support.txt  |  2 +-
>  arch/alpha/Kconfig                                 |  2 +
>  arch/alpha/include/uapi/asm/perf_regs.h            | 45 +++++++++++++++++
>  arch/alpha/kernel/Makefile                         |  1 +
>  arch/alpha/kernel/perf_regs.c                      | 56 ++++++++++++++++++++++
>  6 files changed, 106 insertions(+), 2 deletions(-)
>
> diff --git a/Documentation/features/perf/perf-regs/arch-support.txt b/Documentation/features/perf/perf-regs/arch-support.txt
> index 4c9a5a012075..7d736d678fd4 100644
> --- a/Documentation/features/perf/perf-regs/arch-support.txt
> +++ b/Documentation/features/perf/perf-regs/arch-support.txt
> @@ -6,7 +6,7 @@
>      -----------------------
>      |         arch |status|
>      -----------------------
> -    |       alpha: | TODO |
> +    |       alpha: |  ok  |
>      |         arc: | TODO |
>      |         arm: |  ok  |
>      |       arm64: |  ok  |
> diff --git a/Documentation/features/perf/perf-stackdump/arch-support.txt b/Documentation/features/perf/perf-stackdump/arch-support.txt
> index dd362b5cb638..17b84ea7769d 100644
> --- a/Documentation/features/perf/perf-stackdump/arch-support.txt
> +++ b/Documentation/features/perf/perf-stackdump/arch-support.txt
> @@ -6,7 +6,7 @@
>      -----------------------
>      |         arch |status|
>      -----------------------
> -    |       alpha: | TODO |
> +    |       alpha: |  ok  |
>      |         arc: | TODO |
>      |         arm: |  ok  |
>      |       arm64: |  ok  |
> diff --git a/arch/alpha/Kconfig b/arch/alpha/Kconfig
> index b2c359b947f6..8fafbfabd1e5 100644
> --- a/arch/alpha/Kconfig
> +++ b/arch/alpha/Kconfig
> @@ -41,6 +41,8 @@ config ALPHA
>         select HAVE_PAGE_SIZE_8KB
>         select HAVE_PCSPKR_PLATFORM
>         select HAVE_PERF_EVENTS
> +       select HAVE_PERF_REGS
> +       select HAVE_PERF_USER_STACK_DUMP
>         select NEED_DMA_MAP_STATE
>         select NEED_SG_DMA_LENGTH
>         select GENERIC_IRQ_PROBE
> diff --git a/arch/alpha/include/uapi/asm/perf_regs.h b/arch/alpha/include/uapi/asm/perf_regs.h
> new file mode 100644
> index 000000000000..da96a31b12fb
> --- /dev/null
> +++ b/arch/alpha/include/uapi/asm/perf_regs.h
> @@ -0,0 +1,45 @@
> +/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
> +#ifndef _UAPI_ASM_ALPHA_PERF_REGS_H
> +#define _UAPI_ASM_ALPHA_PERF_REGS_H
> +
> +/*
> + * Indices 0-30 match the hardware (and DWARF) register numbers.
> + * r9-r15 are not saved in pt_regs and cannot be sampled.
> + */
> +enum perf_event_alpha_regs {
> +       PERF_REG_ALPHA_R0,
> +       PERF_REG_ALPHA_R1,
> +       PERF_REG_ALPHA_R2,
> +       PERF_REG_ALPHA_R3,
> +       PERF_REG_ALPHA_R4,
> +       PERF_REG_ALPHA_R5,
> +       PERF_REG_ALPHA_R6,
> +       PERF_REG_ALPHA_R7,
> +       PERF_REG_ALPHA_R8,
> +       PERF_REG_ALPHA_R9,
> +       PERF_REG_ALPHA_R10,
> +       PERF_REG_ALPHA_R11,
> +       PERF_REG_ALPHA_R12,
> +       PERF_REG_ALPHA_R13,
> +       PERF_REG_ALPHA_R14,
> +       PERF_REG_ALPHA_R15,
> +       PERF_REG_ALPHA_R16,
> +       PERF_REG_ALPHA_R17,
> +       PERF_REG_ALPHA_R18,
> +       PERF_REG_ALPHA_R19,
> +       PERF_REG_ALPHA_R20,
> +       PERF_REG_ALPHA_R21,
> +       PERF_REG_ALPHA_R22,
> +       PERF_REG_ALPHA_R23,
> +       PERF_REG_ALPHA_R24,
> +       PERF_REG_ALPHA_R25,
> +       PERF_REG_ALPHA_R26,
> +       PERF_REG_ALPHA_R27,
> +       PERF_REG_ALPHA_R28,
> +       PERF_REG_ALPHA_GP,
> +       PERF_REG_ALPHA_SP,
> +       PERF_REG_ALPHA_PC,
> +       PERF_REG_ALPHA_MAX,
> +};
> +
> +#endif /* _UAPI_ASM_ALPHA_PERF_REGS_H */
> diff --git a/arch/alpha/kernel/Makefile b/arch/alpha/kernel/Makefile
> index 4ea5c189e60e..0c04b89676ed 100644
> --- a/arch/alpha/kernel/Makefile
> +++ b/arch/alpha/kernel/Makefile
> @@ -18,6 +18,7 @@ obj-$(CONFIG_PCI)     += pci.o pci_iommu.o pci-sysfs.o
>  obj-$(CONFIG_SRM_ENV)  += srm_env.o
>  obj-$(CONFIG_MODULES)  += module.o
>  obj-$(CONFIG_PERF_EVENTS) += perf_event.o
> +obj-$(CONFIG_HAVE_PERF_REGS) += perf_regs.o
>  obj-$(CONFIG_RTC_DRV_ALPHA) += rtc.o
>  obj-$(CONFIG_AUDIT)    += audit.o
>
> diff --git a/arch/alpha/kernel/perf_regs.c b/arch/alpha/kernel/perf_regs.c
> new file mode 100644
> index 000000000000..10afe4e205e9
> --- /dev/null
> +++ b/arch/alpha/kernel/perf_regs.c
> @@ -0,0 +1,56 @@
> +// SPDX-License-Identifier: GPL-2.0
> +
> +#include <linux/bits.h>
> +#include <linux/perf_event.h>
> +#include <asm/perf_regs.h>
> +
> +#define REG(idx, field) \
> +       [PERF_REG_ALPHA_##idx] = offsetof(struct pt_regs, field)
> +
> +#define REG_RESERVED   (~GENMASK_ULL(PERF_REG_ALPHA_MAX - 1, 0) | \
> +                        GENMASK_ULL(PERF_REG_ALPHA_R15, PERF_REG_ALPHA_R9))
> +
> +static const unsigned int pt_regs_offset[PERF_REG_ALPHA_MAX] = {
> +       REG(R0, r0),    REG(R1, r1),    REG(R2, r2),    REG(R3, r3),
> +       REG(R4, r4),    REG(R5, r5),    REG(R6, r6),    REG(R7, r7),
> +       REG(R8, r8),    REG(R16, r16),  REG(R17, r17),  REG(R18, r18),
> +       REG(R19, r19),  REG(R20, r20),  REG(R21, r21),  REG(R22, r22),
> +       REG(R23, r23),  REG(R24, r24),  REG(R25, r25),  REG(R26, r26),
> +       REG(R27, r27),  REG(R28, r28),  REG(GP, gp),    REG(PC, pc),
> +};
> +
> +u64 perf_reg_value(struct pt_regs *regs, int idx)
> +{
> +       if (idx == PERF_REG_ALPHA_SP) {
> +               if (user_mode(regs))
> +                       return user_stack_pointer(regs);
> +               /* Same kernel sp approximation as stacktrace.c. */
> +               return (unsigned long)(regs + 1);
> +       }
> +

Tested on my ES40: register masks, SP/PC sampling and user stack dumps
passed with software-clock, page-fault and hardware-cycle events.

For context-switch events, however, regs points to the per-CPU
__perf_regs[0], not an on-stack trap frame. This therefore returns
a per-CPU address as SP.

I reproduced this with:

  perf record -e context-switches -c 1 -Isp,pc -- sleep 1

The sample reported SP=0xfffffc00045b4310 and PC=0. The missing PC
predates this patch, but the incorrect SP is newly exposed here.

One limitation I noticed: for context-switch events, regs points to
the per-CPU synthetic frame, so regs + 1 reports its address rather
than a kernel SP. Alpha already lacks perf_arch_fetch_caller_regs(),
so perhaps synthetic-event register capture is better handled as a
follow-up. Could we document that limitation here?


> +       if (WARN_ON_ONCE((u32)idx >= PERF_REG_ALPHA_MAX ||
> +                        BIT_ULL(idx) & REG_RESERVED))
> +               return 0;
> +
> +       return *(unsigned long *)((void *)regs + pt_regs_offset[idx]);
> +}
> +
> +int perf_reg_validate(u64 mask)
> +{
> +       if (!mask || mask & REG_RESERVED)
> +               return -EINVAL;
> +
> +       return 0;
> +}
> +
> +u64 perf_reg_abi(struct task_struct *task)
> +{
> +       return PERF_SAMPLE_REGS_ABI_64;
> +}
> +
> +void perf_get_regs_user(struct perf_regs *regs_user,
> +                       struct pt_regs *regs)
> +{
> +       regs_user->regs = task_pt_regs(current);
> +       regs_user->abi = perf_reg_abi(current);
> +}
>
> --
> 2.55.0
>

Tested-by: Magnus Lindholm <linmag7@gmail.com>