The task of ftrace_arch_code_modify(_post)_prepare() caller is
stop_machine, whose caller and work thread are of different tasks. The
lockdep checker needs the same task context, or it's wrong. That means
it's a bug here to use lockdep_assert_held because we don't guarantee
the same task context.
kernel/locking/lockdep.c:
int __lock_is_held(const struct lockdep_map *lock, int read)
{
struct task_struct *curr = current;
int i;
for (i = 0; i < curr->lockdep_depth; i++) {
^^^^^^^^^^^^^^^^^^^
struct held_lock *hlock = curr->held_locks + i;
^^^^^^^^^^^^^^^^
if (match_held_lock(hlock, lock)) {
if (read == -1 || !!hlock->read == read)
return LOCK_STATE_HELD;
The __lock_is_held depends on current held_locks records; if
stop_machine makes the checker running on another task, that's wrong.
Here is the log:
[ 15.761523] ------------[ cut here ]------------
[ 15.762125] WARNING: CPU: 0 PID: 15 at arch/riscv/kernel/patch.c:63 patch_insn_write+0x72/0x364
[ 15.763258] Modules linked in:
[ 15.764154] CPU: 0 PID: 15 Comm: migration/0 Not tainted 6.1.0-rc1-00014-g66924be85884-dirty #377
[ 15.765339] Hardware name: riscv-virtio,qemu (DT)
[ 15.765985] Stopper: multi_cpu_stop+0x0/0x192 <- stop_cpus.constprop.0+0x90/0xe2
[ 15.766711] epc : patch_insn_write+0x72/0x364
[ 15.767011] ra : patch_insn_write+0x70/0x364
[ 15.767276] epc : ffffffff8000721e ra : ffffffff8000721c sp : ff2000000067bca0
[ 15.767622] gp : ffffffff81603f90 tp : ff60000002432a00 t0 : 7300000000000000
[ 15.767919] t1 : 0000000000000000 t2 : 73695f6b636f6c5f s0 : ff2000000067bcf0
[ 15.768238] s1 : 0000000000000008 a0 : 0000000000000000 a1 : 0000000000000000
[ 15.768537] a2 : 0000000000000000 a3 : 0000000000000000 a4 : 0000000000000000
[ 15.768837] a5 : 0000000000000000 a6 : 0000000000000000 a7 : 0000000000000000
[ 15.769139] s2 : ffffffff80009faa s3 : ff2000000067bd10 s4 : ffffffffffffffff
[ 15.769447] s5 : 0000000000000001 s6 : 0000000000000001 s7 : 0000000000000003
[ 15.769740] s8 : 0000000000000002 s9 : 0000000000000004 s10: 0000000000000003
[ 15.770027] s11: 0000000000000002 t3 : 0000000000000000 t4 : ffffffff819af097
[ 15.770323] t5 : ffffffff819af098 t6 : ff2000000067ba28
[ 15.770574] status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
[ 15.771102] [<ffffffff80007520>] patch_text_nosync+0x10/0x3a
[ 15.771421] [<ffffffff80009c66>] ftrace_update_ftrace_func+0x74/0x10a
[ 15.771704] [<ffffffff800fa17e>] ftrace_modify_all_code+0xb0/0x16c
[ 15.771958] [<ffffffff800fa24c>] __ftrace_modify_code+0x12/0x1c
[ 15.772196] [<ffffffff800e110e>] multi_cpu_stop+0x14a/0x192
[ 15.772454] [<ffffffff800e0a34>] cpu_stopper_thread+0x96/0x14c
[ 15.772699] [<ffffffff8003f4ea>] smpboot_thread_fn+0xf8/0x1cc
[ 15.772945] [<ffffffff8003ac9c>] kthread+0xe2/0xf8
[ 15.773160] [<ffffffff80003e98>] ret_from_exception+0x0/0x14
[ 15.773471] ---[ end trace 0000000000000000 ]---
By the way, this also fixes the same issue for patch_text().
Fixes: 0ff7c3b33127 ("riscv: Use text_mutex instead of patch_lock")
Co-developed-by: Guo Ren <[email protected]>
Signed-off-by: Guo Ren <[email protected]>
Cc: Zong Li <[email protected]>
Cc: Palmer Dabbelt <[email protected]>
Signed-off-by: Changbin Du <[email protected]>
---
Changes in v4:
- preserve and update comments.
Changes in v3:
- denote this also fixes function patch_text().
Changes in v2:
- Rewrite commit log with lockdep explanation [Guo Ren]
- Rebase on v6.1 [Guo Ren]
v1:
https://lore.kernel.org/linux-riscv/[email protected]/
---
arch/riscv/kernel/ftrace.c | 5 ++---
arch/riscv/kernel/patch.c | 15 ++++++++-------
2 files changed, 10 insertions(+), 10 deletions(-)
diff --git a/arch/riscv/kernel/ftrace.c b/arch/riscv/kernel/ftrace.c
index 2086f6585773..f73660e73822 100644
--- a/arch/riscv/kernel/ftrace.c
+++ b/arch/riscv/kernel/ftrace.c
@@ -126,9 +126,8 @@ int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
/*
* This is called early on, and isn't wrapped by
* ftrace_arch_code_modify_{prepare,post_process}() and therefor doesn't hold
- * text_mutex, which triggers a lockdep failure. SMP isn't running so we could
- * just directly poke the text, but it's simpler to just take the lock
- * ourselves.
+ * text_mutex. SMP isn't running so we could just directly poke the text, but
+ * it's simpler to just take the lock ourselves.
*/
int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec)
{
diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c
index 765004b60513..8eb243703efe 100644
--- a/arch/riscv/kernel/patch.c
+++ b/arch/riscv/kernel/patch.c
@@ -49,19 +49,20 @@ static void patch_unmap(int fixmap)
}
NOKPROBE_SYMBOL(patch_unmap);
+/*
+ * Before reaching here, it was expected to lock the text_mutex
+ * already, so we don't need to give another lock here and could
+ * ensure that it was safe between each cores. We do not add
+ * lockdep assertion here since it would trigger a false positive
+ * when called by stop_machine (The lockdep checker requires the
+ * same task context).
+ */
static int patch_insn_write(void *addr, const void *insn, size_t len)
{
void *waddr = addr;
bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
int ret;
- /*
- * Before reaching here, it was expected to lock the text_mutex
- * already, so we don't need to give another lock here and could
- * ensure that it was safe between each cores.
- */
- lockdep_assert_held(&text_mutex);
-
if (across_pages)
patch_map(addr + len, FIX_TEXT_POKE1);
--
2.25.1
Hey Changbin,
On Thu, Feb 02, 2023 at 07:41:16PM +0800, Changbin Du wrote:
> The task of ftrace_arch_code_modify(_post)_prepare() caller is
> stop_machine, whose caller and work thread are of different tasks. The
> lockdep checker needs the same task context, or it's wrong. That means
> it's a bug here to use lockdep_assert_held because we don't guarantee
> the same task context.
>
> kernel/locking/lockdep.c:
> int __lock_is_held(const struct lockdep_map *lock, int read)
> {
> struct task_struct *curr = current;
> int i;
>
> for (i = 0; i < curr->lockdep_depth; i++) {
> ^^^^^^^^^^^^^^^^^^^
> struct held_lock *hlock = curr->held_locks + i;
> ^^^^^^^^^^^^^^^^
> if (match_held_lock(hlock, lock)) {
> if (read == -1 || !!hlock->read == read)
> return LOCK_STATE_HELD;
>
> The __lock_is_held depends on current held_locks records; if
> stop_machine makes the checker running on another task, that's wrong.
>
> Here is the log:
> [ 15.761523] ------------[ cut here ]------------
> [ 15.762125] WARNING: CPU: 0 PID: 15 at arch/riscv/kernel/patch.c:63 patch_insn_write+0x72/0x364
> [ 15.763258] Modules linked in:
> [ 15.764154] CPU: 0 PID: 15 Comm: migration/0 Not tainted 6.1.0-rc1-00014-g66924be85884-dirty #377
> [ 15.765339] Hardware name: riscv-virtio,qemu (DT)
> [ 15.765985] Stopper: multi_cpu_stop+0x0/0x192 <- stop_cpus.constprop.0+0x90/0xe2
> [ 15.766711] epc : patch_insn_write+0x72/0x364
> [ 15.767011] ra : patch_insn_write+0x70/0x364
> [ 15.767276] epc : ffffffff8000721e ra : ffffffff8000721c sp : ff2000000067bca0
> [ 15.767622] gp : ffffffff81603f90 tp : ff60000002432a00 t0 : 7300000000000000
> [ 15.767919] t1 : 0000000000000000 t2 : 73695f6b636f6c5f s0 : ff2000000067bcf0
> [ 15.768238] s1 : 0000000000000008 a0 : 0000000000000000 a1 : 0000000000000000
> [ 15.768537] a2 : 0000000000000000 a3 : 0000000000000000 a4 : 0000000000000000
> [ 15.768837] a5 : 0000000000000000 a6 : 0000000000000000 a7 : 0000000000000000
> [ 15.769139] s2 : ffffffff80009faa s3 : ff2000000067bd10 s4 : ffffffffffffffff
> [ 15.769447] s5 : 0000000000000001 s6 : 0000000000000001 s7 : 0000000000000003
> [ 15.769740] s8 : 0000000000000002 s9 : 0000000000000004 s10: 0000000000000003
> [ 15.770027] s11: 0000000000000002 t3 : 0000000000000000 t4 : ffffffff819af097
> [ 15.770323] t5 : ffffffff819af098 t6 : ff2000000067ba28
> [ 15.770574] status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
> [ 15.771102] [<ffffffff80007520>] patch_text_nosync+0x10/0x3a
> [ 15.771421] [<ffffffff80009c66>] ftrace_update_ftrace_func+0x74/0x10a
> [ 15.771704] [<ffffffff800fa17e>] ftrace_modify_all_code+0xb0/0x16c
> [ 15.771958] [<ffffffff800fa24c>] __ftrace_modify_code+0x12/0x1c
> [ 15.772196] [<ffffffff800e110e>] multi_cpu_stop+0x14a/0x192
> [ 15.772454] [<ffffffff800e0a34>] cpu_stopper_thread+0x96/0x14c
> [ 15.772699] [<ffffffff8003f4ea>] smpboot_thread_fn+0xf8/0x1cc
> [ 15.772945] [<ffffffff8003ac9c>] kthread+0xe2/0xf8
> [ 15.773160] [<ffffffff80003e98>] ret_from_exception+0x0/0x14
> [ 15.773471] ---[ end trace 0000000000000000 ]---
>
> By the way, this also fixes the same issue for patch_text().
Given this lockdep stuff seems to have pointed out that we weren't
taking the lock for alternative patching just this past week [1], I'm
really not convinced that deleting this is a good idea.
Thanks,
Conor.
1 - https://patchwork.kernel.org/project/linux-riscv/patch/[email protected]/
>
> Fixes: 0ff7c3b33127 ("riscv: Use text_mutex instead of patch_lock")
> Co-developed-by: Guo Ren <[email protected]>
> Signed-off-by: Guo Ren <[email protected]>
> Cc: Zong Li <[email protected]>
> Cc: Palmer Dabbelt <[email protected]>
> Signed-off-by: Changbin Du <[email protected]>
> ---
> Changes in v4:
> - preserve and update comments.
>
> Changes in v3:
> - denote this also fixes function patch_text().
>
> Changes in v2:
> - Rewrite commit log with lockdep explanation [Guo Ren]
> - Rebase on v6.1 [Guo Ren]
>
> v1:
> https://lore.kernel.org/linux-riscv/[email protected]/
> ---
> arch/riscv/kernel/ftrace.c | 5 ++---
> arch/riscv/kernel/patch.c | 15 ++++++++-------
> 2 files changed, 10 insertions(+), 10 deletions(-)
>
> diff --git a/arch/riscv/kernel/ftrace.c b/arch/riscv/kernel/ftrace.c
> index 2086f6585773..f73660e73822 100644
> --- a/arch/riscv/kernel/ftrace.c
> +++ b/arch/riscv/kernel/ftrace.c
> @@ -126,9 +126,8 @@ int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
> /*
> * This is called early on, and isn't wrapped by
> * ftrace_arch_code_modify_{prepare,post_process}() and therefor doesn't hold
> - * text_mutex, which triggers a lockdep failure. SMP isn't running so we could
> - * just directly poke the text, but it's simpler to just take the lock
> - * ourselves.
> + * text_mutex. SMP isn't running so we could just directly poke the text, but
> + * it's simpler to just take the lock ourselves.
> */
> int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec)
> {
> diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c
> index 765004b60513..8eb243703efe 100644
> --- a/arch/riscv/kernel/patch.c
> +++ b/arch/riscv/kernel/patch.c
> @@ -49,19 +49,20 @@ static void patch_unmap(int fixmap)
> }
> NOKPROBE_SYMBOL(patch_unmap);
>
> +/*
> + * Before reaching here, it was expected to lock the text_mutex
> + * already, so we don't need to give another lock here and could
> + * ensure that it was safe between each cores. We do not add
> + * lockdep assertion here since it would trigger a false positive
> + * when called by stop_machine (The lockdep checker requires the
> + * same task context).
> + */
> static int patch_insn_write(void *addr, const void *insn, size_t len)
> {
> void *waddr = addr;
> bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
> int ret;
>
> - /*
> - * Before reaching here, it was expected to lock the text_mutex
> - * already, so we don't need to give another lock here and could
> - * ensure that it was safe between each cores.
> - */
> - lockdep_assert_held(&text_mutex);
> -
> if (across_pages)
> patch_map(addr + len, FIX_TEXT_POKE1);
>
> --
> 2.25.1
>
On Wed, Feb 15, 2023 at 8:22 AM Conor Dooley <[email protected]> wrote:
>
> Hey Changbin,
>
> On Thu, Feb 02, 2023 at 07:41:16PM +0800, Changbin Du wrote:
> > The task of ftrace_arch_code_modify(_post)_prepare() caller is
> > stop_machine, whose caller and work thread are of different tasks. The
> > lockdep checker needs the same task context, or it's wrong. That means
> > it's a bug here to use lockdep_assert_held because we don't guarantee
> > the same task context.
I'm trying to delete all stop_machine in riscv, from ftrace to kprobe.
When I have done, we needn't this patch.
> >
> > kernel/locking/lockdep.c:
> > int __lock_is_held(const struct lockdep_map *lock, int read)
> > {
> > struct task_struct *curr = current;
> > int i;
> >
> > for (i = 0; i < curr->lockdep_depth; i++) {
> > ^^^^^^^^^^^^^^^^^^^
> > struct held_lock *hlock = curr->held_locks + i;
> > ^^^^^^^^^^^^^^^^
> > if (match_held_lock(hlock, lock)) {
> > if (read == -1 || !!hlock->read == read)
> > return LOCK_STATE_HELD;
> >
> > The __lock_is_held depends on current held_locks records; if
> > stop_machine makes the checker running on another task, that's wrong.
> >
> > Here is the log:
> > [ 15.761523] ------------[ cut here ]------------
> > [ 15.762125] WARNING: CPU: 0 PID: 15 at arch/riscv/kernel/patch.c:63 patch_insn_write+0x72/0x364
> > [ 15.763258] Modules linked in:
> > [ 15.764154] CPU: 0 PID: 15 Comm: migration/0 Not tainted 6.1.0-rc1-00014-g66924be85884-dirty #377
> > [ 15.765339] Hardware name: riscv-virtio,qemu (DT)
> > [ 15.765985] Stopper: multi_cpu_stop+0x0/0x192 <- stop_cpus.constprop.0+0x90/0xe2
> > [ 15.766711] epc : patch_insn_write+0x72/0x364
> > [ 15.767011] ra : patch_insn_write+0x70/0x364
> > [ 15.767276] epc : ffffffff8000721e ra : ffffffff8000721c sp : ff2000000067bca0
> > [ 15.767622] gp : ffffffff81603f90 tp : ff60000002432a00 t0 : 7300000000000000
> > [ 15.767919] t1 : 0000000000000000 t2 : 73695f6b636f6c5f s0 : ff2000000067bcf0
> > [ 15.768238] s1 : 0000000000000008 a0 : 0000000000000000 a1 : 0000000000000000
> > [ 15.768537] a2 : 0000000000000000 a3 : 0000000000000000 a4 : 0000000000000000
> > [ 15.768837] a5 : 0000000000000000 a6 : 0000000000000000 a7 : 0000000000000000
> > [ 15.769139] s2 : ffffffff80009faa s3 : ff2000000067bd10 s4 : ffffffffffffffff
> > [ 15.769447] s5 : 0000000000000001 s6 : 0000000000000001 s7 : 0000000000000003
> > [ 15.769740] s8 : 0000000000000002 s9 : 0000000000000004 s10: 0000000000000003
> > [ 15.770027] s11: 0000000000000002 t3 : 0000000000000000 t4 : ffffffff819af097
> > [ 15.770323] t5 : ffffffff819af098 t6 : ff2000000067ba28
> > [ 15.770574] status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
> > [ 15.771102] [<ffffffff80007520>] patch_text_nosync+0x10/0x3a
> > [ 15.771421] [<ffffffff80009c66>] ftrace_update_ftrace_func+0x74/0x10a
> > [ 15.771704] [<ffffffff800fa17e>] ftrace_modify_all_code+0xb0/0x16c
> > [ 15.771958] [<ffffffff800fa24c>] __ftrace_modify_code+0x12/0x1c
> > [ 15.772196] [<ffffffff800e110e>] multi_cpu_stop+0x14a/0x192
> > [ 15.772454] [<ffffffff800e0a34>] cpu_stopper_thread+0x96/0x14c
> > [ 15.772699] [<ffffffff8003f4ea>] smpboot_thread_fn+0xf8/0x1cc
> > [ 15.772945] [<ffffffff8003ac9c>] kthread+0xe2/0xf8
> > [ 15.773160] [<ffffffff80003e98>] ret_from_exception+0x0/0x14
> > [ 15.773471] ---[ end trace 0000000000000000 ]---
> >
> > By the way, this also fixes the same issue for patch_text().
>
> Given this lockdep stuff seems to have pointed out that we weren't
> taking the lock for alternative patching just this past week [1], I'm
> really not convinced that deleting this is a good idea.
>
> Thanks,
> Conor.
>
> 1 - https://patchwork.kernel.org/project/linux-riscv/patch/[email protected]/
>
> >
> > Fixes: 0ff7c3b33127 ("riscv: Use text_mutex instead of patch_lock")
> > Co-developed-by: Guo Ren <[email protected]>
> > Signed-off-by: Guo Ren <[email protected]>
> > Cc: Zong Li <[email protected]>
> > Cc: Palmer Dabbelt <[email protected]>
> > Signed-off-by: Changbin Du <[email protected]>
> > ---
> > Changes in v4:
> > - preserve and update comments.
> >
> > Changes in v3:
> > - denote this also fixes function patch_text().
> >
> > Changes in v2:
> > - Rewrite commit log with lockdep explanation [Guo Ren]
> > - Rebase on v6.1 [Guo Ren]
> >
> > v1:
> > https://lore.kernel.org/linux-riscv/[email protected]/
> > ---
> > arch/riscv/kernel/ftrace.c | 5 ++---
> > arch/riscv/kernel/patch.c | 15 ++++++++-------
> > 2 files changed, 10 insertions(+), 10 deletions(-)
> >
> > diff --git a/arch/riscv/kernel/ftrace.c b/arch/riscv/kernel/ftrace.c
> > index 2086f6585773..f73660e73822 100644
> > --- a/arch/riscv/kernel/ftrace.c
> > +++ b/arch/riscv/kernel/ftrace.c
> > @@ -126,9 +126,8 @@ int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
> > /*
> > * This is called early on, and isn't wrapped by
> > * ftrace_arch_code_modify_{prepare,post_process}() and therefor doesn't hold
> > - * text_mutex, which triggers a lockdep failure. SMP isn't running so we could
> > - * just directly poke the text, but it's simpler to just take the lock
> > - * ourselves.
> > + * text_mutex. SMP isn't running so we could just directly poke the text, but
> > + * it's simpler to just take the lock ourselves.
> > */
> > int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec)
> > {
> > diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c
> > index 765004b60513..8eb243703efe 100644
> > --- a/arch/riscv/kernel/patch.c
> > +++ b/arch/riscv/kernel/patch.c
> > @@ -49,19 +49,20 @@ static void patch_unmap(int fixmap)
> > }
> > NOKPROBE_SYMBOL(patch_unmap);
> >
> > +/*
> > + * Before reaching here, it was expected to lock the text_mutex
> > + * already, so we don't need to give another lock here and could
> > + * ensure that it was safe between each cores. We do not add
> > + * lockdep assertion here since it would trigger a false positive
> > + * when called by stop_machine (The lockdep checker requires the
> > + * same task context).
> > + */
> > static int patch_insn_write(void *addr, const void *insn, size_t len)
> > {
> > void *waddr = addr;
> > bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
> > int ret;
> >
> > - /*
> > - * Before reaching here, it was expected to lock the text_mutex
> > - * already, so we don't need to give another lock here and could
> > - * ensure that it was safe between each cores.
> > - */
> > - lockdep_assert_held(&text_mutex);
> > -
> > if (across_pages)
> > patch_map(addr + len, FIX_TEXT_POKE1);
> >
> > --
> > 2.25.1
> >
--
Best Regards
Guo Ren
On Wed, Feb 15, 2023 at 09:24:33AM +0800, Guo Ren wrote:
> On Wed, Feb 15, 2023 at 8:22 AM Conor Dooley <[email protected]> wrote:
> >
> > Hey Changbin,
> >
> > On Thu, Feb 02, 2023 at 07:41:16PM +0800, Changbin Du wrote:
> > > The task of ftrace_arch_code_modify(_post)_prepare() caller is
> > > stop_machine, whose caller and work thread are of different tasks. The
> > > lockdep checker needs the same task context, or it's wrong. That means
> > > it's a bug here to use lockdep_assert_held because we don't guarantee
> > > the same task context.
> I'm trying to delete all stop_machine in riscv, from ftrace to kprobe.
> When I have done, we needn't this patch.
>
Which approch would you use? I looked through the riscv-spec, but didn't find any
description abount concurrent modification and execution.
> > >
> > > 2.25.1
> > >
>
>
>
> --
> Best Regards
> Guo Ren
--
Cheers,
Changbin Du
Changbin Du <[email protected]> writes:
> On Wed, Feb 15, 2023 at 09:24:33AM +0800, Guo Ren wrote:
>> On Wed, Feb 15, 2023 at 8:22 AM Conor Dooley <[email protected]> wrote:
>> >
>> > Hey Changbin,
>> >
>> > On Thu, Feb 02, 2023 at 07:41:16PM +0800, Changbin Du wrote:
>> > > The task of ftrace_arch_code_modify(_post)_prepare() caller is
>> > > stop_machine, whose caller and work thread are of different tasks. The
>> > > lockdep checker needs the same task context, or it's wrong. That means
>> > > it's a bug here to use lockdep_assert_held because we don't guarantee
>> > > the same task context.
>> I'm trying to delete all stop_machine in riscv, from ftrace to kprobe.
>> When I have done, we needn't this patch.
>>
> Which approch would you use? I looked through the riscv-spec, but didn't find any
> description abount concurrent modification and execution.
CMODX is not specified for RISC-V yet, unfortunately.
This has been discussed here [1]. Maybe we can start with stating for
which implementations Guo's approach work?
[1] https://lore.kernel.org/all/CAJF2gTS0s4X_uwLaEeSqKAyRmxCR2vxRuHhz7-SP2w4bBqzr+Q@mail.gmail.com/