Skip to contentSkip to conversation

tools/mm: add hwpoison-panic tool

PATCH Superseded Breno Leitao 4 participants 5 messages Aug 3 – Sep 7

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 tools/mm: add hwpoison-panic tool

Add a tool that enables the vm.panic_on_unrecoverable_memory_failure sysctl, picks a kernel-owned PFN and writes its physical address to hard_offline_page. Three page kinds are selectable with -k: rodata (default), slab or pgtable. In all cases the host should panic.

Example:

        # ./hwpoison-panic -k slab --yes-panic-my-kernel
        injecting hwpoison at phys 0x100032000 (pfn 0x100032, kind=slab)
        expecting kernel panic: 'Memory failure: <pfn>: unrecoverable page'

In dmesg, you will see:

        Memory failure: 0x100032: unhandlable page.
        Memory failure: 0x100032: recovery action for reserved kernel page: Ignored
        Kernel panic - not syncing: Memory failure: 0x100032: unrecoverable page

This lives in tools/mm rather than selftests/mm because every successful run crashes the machine, which is not something to run from CI.

The --yes-panic-my-kernel argument is required so an accidental invocation does not take the box down.

---
Changes in v2:
- Reword the file header, the tool triggers a panic rather than
confirming one (SJ Park)
- Rename -f to --yes-panic-my-kernel and say in usage() what it costs
(SJ Park)
- Call check_prereqs() before getpagesize() (SJ Park)
- Use 64-bit types for physical addresses and PFNs. On a 32-bit build
the old long/unsigned long pair truncated addresses above 4G, and the
negative error sentinel was indistinguishable from a valid address at
or above 2G
- Link to v1: https://patch.msgid.link/20260731-memory_failure_rewrite_test-v1-1-6aa8c6435693@debian.org

To: Andrew Morton <[email protected]>
To: Miaohe Lin <[email protected]>
To: Naoya Horiguchi <[email protected]>
To: Paul Walmsley <[email protected]>
To: Palmer Dabbelt <[email protected]>
To: Albert Ou <[email protected]>
To: Alexandre Ghiti <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]

---
 MAINTAINERS                              |   1 +
 tools/mm/.gitignore                      |   1 +
 tools/mm/Makefile                        |   2 +
 tools/mm/memory-failure/hwpoison-panic.c | 358 +++++++++++++++++++++++++++++++
 4 files changed, 362 insertions(+)
MAINTAINERS +1 -0
index 56b337a064478..799c4883d5b0b 100644
@@ -12094,6 +12094,7 @@ F:	include/linux/memory-failure.h
 F:	include/trace/events/memory-failure.h
 F:	mm/hwpoison-inject.c
 F:	mm/memory-failure.c
+F:	tools/mm/memory-failure/
 F:	tools/testing/selftests/mm/memory-failure.c
 
 HYCON HY46XX TOUCHSCREEN SUPPORT
tools/mm/.gitignore +1 -0
index 1446a659e5408..b9b80d909db4a 100644
@@ -3,3 +3,4 @@ slabinfo
 page-types
 page_owner_sort
 thp_swap_allocator_test
+memory-failure/hwpoison-panic
tools/mm/Makefile +2 -0
index 858186a6eefdb..9f6de957a95a0 100644
@@ -4,6 +4,7 @@
 include ../scripts/Makefile.include
 
 BUILD_TARGETS=page-types slabinfo page_owner_sort page_owner_filter thp_swap_allocator_test
+BUILD_TARGETS+=memory-failure/hwpoison-panic
 INSTALL_TARGETS = $(BUILD_TARGETS) thpmaps
 
 LIB_DIR = ../lib/api
@@ -24,6 +25,7 @@ $(LIBS):
 
 clean:
 	$(RM) page-types slabinfo page_owner_sort page_owner_filter thp_swap_allocator_test
+	$(RM) memory-failure/hwpoison-panic
 	make -C $(LIB_DIR) clean
 
 sbindir ?= /usr/sbin
tools/mm/memory-failure/hwpoison-panic.c new file +358 -0

358 lines not on this page · raw message

---
base-commit: 95d6a9ccef99117115e41e9adb271243bd5e985b
change-id: 20260728-memory_failure_rewrite_test-7a309a601a6e

Best regards,

Signature · 1 line
Breno Leitao <[email protected]>
#1 lore raw
Quoted 17 lines · Breno Leitao

Add a tool that enables the vm.panic_on_unrecoverable_memory_failure sysctl, picks a kernel-owned PFN and writes its physical address to hard_offline_page. Three page kinds are selectable with -k: rodata (default), slab or pgtable. In all cases the host should panic.

Example:

        # ./hwpoison-panic -k slab --yes-panic-my-kernel
        injecting hwpoison at phys 0x100032000 (pfn 0x100032, kind=slab)
        expecting kernel panic: 'Memory failure: <pfn>: unrecoverable page'

In dmesg, you will see:

        Memory failure: 0x100032: unhandlable page.
        Memory failure: 0x100032: recovery action for reserved kernel page: Ignored
        Kernel panic - not syncing: Memory failure: 0x100032: unrecoverable page
Quoted 7 lines

This lives in tools/mm rather than selftests/mm because every successful run crashes the machine, which is not something to run from CI.

The --yes-panic-my-kernel argument is required so an accidental invocation does not take the box down.

This patch looks good to me with some nits below.

Quoted 12 lines

---
Changes in v2:
- Reword the file header, the tool triggers a panic rather than
confirming one (SJ Park)
- Rename -f to --yes-panic-my-kernel and say in usage() what it costs
(SJ Park)
- Call check_prereqs() before getpagesize() (SJ Park)
- Use 64-bit types for physical addresses and PFNs. On a 32-bit build
the old long/unsigned long pair truncated addresses above 4G, and the
negative error sentinel was indistinguishable from a valid address at
or above 2G
- Link to v1: https://patch.msgid.link/20260731-memory_failure_rewrite_test-v1-1-6aa8c6435693@debian.org

...

tools/mm/memory-failure/hwpoison-panic.c line 119

Quoted 4 lines
tools/mm/memory-failure/hwpoison-panic.c new file +4 -0
+
+static int pick_kpageflags_phys_addr(uint64_t want, uint64_t *phys_addr)
+{
+	uint64_t pfn = (16UL << 20) / page_size;

Could we use SZ_16M macro here?

Quoted 45 lines
tools/mm/memory-failure/hwpoison-panic.c new file +45 -0
+	uint64_t flags;
+	int ret = -1;
+	int fd;
+
+	fd = open(PROC_KPAGEFLAGS, O_RDONLY);
+	if (fd < 0)
+		return -1;
+
+	for (; kpageflags_read(fd, pfn, &flags) == 0; pfn++) {
+		if ((flags & want) && !(flags & BIT(HWPOISON)) &&
+		    !(flags & BIT(NOPAGE)) && !(flags & BIT(COMPOUND_TAIL))) {
+			*phys_addr = pfn * page_size;
+			ret = 0;
+			break;
+		}
+	}
+
+	close(fd);
+	return ret;
+}
+
+static int read_sysctl(unsigned long *val)
+{
+	FILE *f = fopen(SYSCTL_PATH, "r");
+	int ret;
+
+	if (!f)
+		return -1;
+	ret = fscanf(f, "%lu", val) == 1 ? 0 : -1;
+	fclose(f);
+
+	return ret;
+}
+
+static int write_sysctl(unsigned long val)
+{
+	FILE *f = fopen(SYSCTL_PATH, "w");
+	int ret;
+
+	if (!f)
+		return -1;
+	ret = fprintf(f, "%lu", val) < 0 ? -1 : 0;
+	fclose(f);
+
+	return ret;

tools/mm/memory-failure/hwpoison-panic.c line 174

Quoted 10 lines
tools/mm/memory-failure/hwpoison-panic.c new file +10 -0
+}
+
+/* hard_offline_page() injects with MF_SW_SIMULATED, so unpoison is allowed. */
+static void unpoison_pfn(uint64_t pfn)
+{
+	char path[PATH_MAX], buf[32];
+	const char *debugfs;
+	int fd, len;
+
+	debugfs = debugfs__mount();

I might be miss something but I can't find the implementation of debugfs__mount.

Quoted 14 lines
tools/mm/memory-failure/hwpoison-panic.c new file +14 -0
+	if (!debugfs)
+		return;
+
+	snprintf(path, sizeof(path), "%s/hwpoison/unpoison-pfn", debugfs);
+	fd = open(path, O_WRONLY);
+	if (fd < 0)
+		return;
+
+	len = snprintf(buf, sizeof(buf), "0x%llx\n", (unsigned long long)pfn);
+	if (write(fd, buf, len) < 0)
+		perror("unpoison-pfn");
+	close(fd);
+}
+

...

tools/mm/memory-failure/hwpoison-panic.c line 283

Quoted 10 lines
tools/mm/memory-failure/hwpoison-panic.c new file +10 -0
+
+static void check_prereqs(int armed)
+{
+	if (geteuid())
+		fatal("must run as root\n");
+	if (access(SYSCTL_PATH, W_OK))
+		fatal("%s not present (kernel without the sysctl?)\n",
+		      SYSCTL_PATH);
+	if (access(INJECT_PATH, W_OK))
+		fatal("%s not present (no MEMORY_HOTPLUG?)\n", INJECT_PATH);

MEMORY_HOTPLUG?

Thanks.
.

#2 lore raw

Hello Miaohe,

First of all, thanks for looking at this one.

Quoted 6 lines · Miaohe Lin

tools/mm/memory-failure/hwpoison-panic.c line 119

Quoted 3 lines · Breno Leitao
tools/mm/memory-failure/hwpoison-panic.c new file +3 -0
+static int pick_kpageflags_phys_addr(uint64_t want, uint64_t *phys_addr)
+{
+	uint64_t pfn = (16UL << 20) / page_size;

Could we use SZ_16M macro here?

Sure, for that I will include ../../include/linux/sizes.h in the next version.

Quoted 10 lines

tools/mm/memory-failure/hwpoison-panic.c line 174

Quoted 8 lines
tools/mm/memory-failure/hwpoison-panic.c new file +8 -0
+/* hard_offline_page() injects with MF_SW_SIMULATED, so unpoison is allowed. */
+static void unpoison_pfn(uint64_t pfn)
+{
+	char path[PATH_MAX], buf[32];
+	const char *debugfs;
+	int fd, len;
+
+	debugfs = debugfs__mount();

I might be miss something but I can't find the implementation of debugfs__mount.

It is an ugly macro declared in tools/lib/api/fs/fs.h:

        #define FS(name)                                \
                const char *name##__mountpoint(void);   \
                const char *name##__mount(void);        \
                bool name##__configured(void);

And then:
FS(debugfs)

Quoted 4 lines

tools/mm/memory-failure/hwpoison-panic.c line 283

Quoted 2 lines
tools/mm/memory-failure/hwpoison-panic.c new file +2 -0
+	if (access(INJECT_PATH, W_OK))
+		fatal("%s not present (no MEMORY_HOTPLUG?)\n", INJECT_PATH);

MEMORY_HOTPLUG?

lol. it should be MEMORY_FAILURE

I will get these points fixed and respin.
--breno

maintainer
#3 lore raw
In reply to Breno Leitao
Quoted 11 lines · On Mon, 03 Aug 2026 04:31:20 -0700 Breno Leitao <[email protected]> wrote:

Add a tool that enables the vm.panic_on_unrecoverable_memory_failure sysctl, picks a kernel-owned PFN and writes its physical address to hard_offline_page. Three page kinds are selectable with -k: rodata (default), slab or pgtable. In all cases the host should panic.

Example:

        # ./hwpoison-panic -k slab --yes-panic-my-kernel
        injecting hwpoison at phys 0x100032000 (pfn 0x100032, kind=slab)
        expecting kernel panic: 'Memory failure: <pfn>: unrecoverable page'
Quoted 8 lines

In dmesg, you will see:

        Memory failure: 0x100032: unhandlable page.
        Memory failure: 0x100032: recovery action for reserved kernel page: Ignored
        Kernel panic - not syncing: Memory failure: 0x100032: unrecoverable page

This lives in tools/mm rather than selftests/mm because every successful run crashes the machine, which is not something to run from CI.

I wonder if we can fix this. Add a knob which tells the kernel not to panic/oops/crash/etc. Instead it sends a message "i tried to crash but you told me not to" and proceeds.

Set that knob around your hwpoison-panic tool and voila, it's now a selftest.

Wondering if we can sneakily use fault-injection to do this. Add a new i_should_crash_now() and use fault-injection to make this function "fail", by returning zero. Or something else.

#4 lore raw
Quoted 22 lines · Andrew Morton
Quoted 11 lines · On Mon, 03 Aug 2026 04:31:20 -0700 Breno Leitao <[email protected]> wrote:

Add a tool that enables the vm.panic_on_unrecoverable_memory_failure sysctl, picks a kernel-owned PFN and writes its physical address to hard_offline_page. Three page kinds are selectable with -k: rodata (default), slab or pgtable. In all cases the host should panic.

Example:

        # ./hwpoison-panic -k slab --yes-panic-my-kernel
        injecting hwpoison at phys 0x100032000 (pfn 0x100032, kind=slab)
        expecting kernel panic: 'Memory failure: <pfn>: unrecoverable page'
Quoted 8 lines

In dmesg, you will see:

        Memory failure: 0x100032: unhandlable page.
        Memory failure: 0x100032: recovery action for reserved kernel page: Ignored
        Kernel panic - not syncing: Memory failure: 0x100032: unrecoverable page

This lives in tools/mm rather than selftests/mm because every successful run crashes the machine, which is not something to run from CI.

Quoted 6 lines

I wonder if we can fix this. Add a knob which tells the kernel not to panic/oops/crash/etc. Instead it sends a message "i tried to crash but you told me not to" and proceeds.

Set that knob around your hwpoison-panic tool and voila, it's now a selftest.

We just shot down a random "kernel-owned PFN" (marking it hwpoison etc).

That's really not something you want to do from a selftest.

You'd need some kind of hwpoison simulation mode.

Signature · 3 lines
Cheers,

David