Please have each new patch be a new thread, and not a Cc to the previous version of the patch. As it makes it hard to find in INBOXs.

On Mon, 10 Apr 2023 15:35:08 +0800 Tze-nan Wu Tze-nan.Wu@mediatek.com wrote:

Write to buffer_size_kb can permanently fail, due to cpu_online_mask may changed between two for_each_online_buffer_cpu loops. The number of increasing and decreasing on cpu_buffer->resize_disable may be inconsistent, leading that the resize_disabled in some CPUs becoming none zero after ring_buffer_reset_online_cpus return.

This issue can be reproduced by "echo 0 > trace" while hotplugging cpu. After reproducing success, we can find out buffer_size_kb will not be functional anymore.

Prevent the two "loops" in this function from iterating through different online cpus by copying cpu_online_mask at the entry of the function.

The "Changes from" need to go below the '---', otherwise they are added to the git commit (we don't want it there).

Changes from v1 to v3: Declare the cpumask variable statically rather than dynamically.

Changes from v2 to v3: Considering holding cpu_hotplug_lock too long because of the synchronize_rcu(), maybe it's better to prevent the issue by copying cpu_online_mask at the entry of the function as V1 does, instead of using cpus_read_lock().

Link: https://lore.kernel.org/lkml/20230408052226.25268-1-Tze-nan.Wu@mediatek.com/ Link: https://lore.kernel.org/oe-kbuild-all/202304082051.Dp50upfS-lkp@intel.com/ Link: https://lore.kernel.org/oe-kbuild-all/202304081615.eiaqpbV8-lkp@intel.com/

Cc: stable@vger.kernel.org Cc: npiggin@gmail.com Fixes: b23d7a5f4a07 ("ring-buffer: speed up buffer resets by avoiding synchronize_rcu for each CPU") Reported-by: kernel test robot lkp@intel.com Reviewed-by: Cheng-Jui Wang cheng-jui.wang@mediatek.com Signed-off-by: Tze-nan Wu Tze-nan.Wu@mediatek.com

---

This is where the "Changes from" go. And since this patch is not suppose to be a Cc. But since it's still good to have a link to it. You could do:

Changes from v2 to v3: https://lore.kernel.org/linux-trace-kernel/20230409024616.31099-1-Tze-nan.Wu... Considering holding cpu_hotplug_lock too long because of the synchronize_rcu(), maybe it's better to prevent the issue by copying cpu_online_mask at the entry of the function as V1 does, instead of using cpus_read_lock().

Where the previous version changes has the lore link to the previous patch, in case someone wants to look at it.

kernel/trace/ring_buffer.c | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c index 76a2d91eecad..dc758930dacb 100644 --- a/kernel/trace/ring_buffer.c +++ b/kernel/trace/ring_buffer.c @@ -288,9 +288,6 @@ EXPORT_SYMBOL_GPL(ring_buffer_event_data); #define for_each_buffer_cpu(buffer, cpu) \ for_each_cpu(cpu, buffer->cpumask) -#define for_each_online_buffer_cpu(buffer, cpu) \

• for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)

#define TS_SHIFT 27 #define TS_MASK ((1ULL << TS_SHIFT) - 1) #define TS_DELTA_TEST (~TS_MASK) @@ -5353,12 +5350,19 @@ EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu); void ring_buffer_reset_online_cpus(struct trace_buffer *buffer) { struct ring_buffer_per_cpu *cpu_buffer;

• cpumask_t reset_online_cpumask;

It's usually considered bad form to put a cpumask on the stack. As it can be 128 bytes for a machine with 1024 CPUs (and yes they do exist). Also, the mask size is set to NR_CPUS not the actual size, so you do not even need to have it that big.

int cpu;

• /*

• * Record cpu_online_mask here to make sure we iterate through the same
  

• * online CPUs in the following two loops.
  

• */
  

• cpumask_copy(&reset_online_cpumask, cpu_online_mask);
• /* prevent another thread from changing buffer sizes */ mutex_lock(&buffer->mutex);

• for_each_online_buffer_cpu(buffer, cpu) {

• for_each_cpu_and(cpu, buffer->cpumask, &reset_online_cpumask) { cpu_buffer = buffer->buffers[cpu];

atomic_inc(&cpu_buffer->resize_disabled);

Anyway, we don't need to modify any of the above, and just do the following instead of atomic_inc():

#define RESET_BIT (1 << 30)

atomic_add(&cpu_buffer->resize_disabled, RESET_BIT);

@@ -5368,7 +5372,7 @@ void ring_buffer_reset_online_cpus(struct trace_buffer *buffer) /* Make sure all commits have finished */ synchronize_rcu();

• for_each_online_buffer_cpu(buffer, cpu) {

• for_each_cpu_and(cpu, buffer->cpumask, &reset_online_cpumask) { cpu_buffer = buffer->buffers[cpu];

Then here we can do:

/* * If a CPU came online during the synchronize_rcu(), then * ignore it. */ if (!atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT)) continue;

atomic_sub(&cpu_buffer->resize_disabled, RESET_BIT);

As the resize_disabled only needs to be set to something to make it disabled.

-- Steve

reset_disabled_cpu_buffer(cpu_buffer);

Hi Steve,

On Tue, 2023-04-11 at 12:44 -0400, Steven Rostedt wrote:

External email : Please do not click links or open attachments until you have verified the sender or the content.

Please have each new patch be a new thread, and not a Cc to the previous version of the patch. As it makes it hard to find in INBOXs.

No problem, got it.

On Mon, 10 Apr 2023 15:35:08 +0800 Tze-nan Wu Tze-nan.Wu@mediatek.com wrote:

...

Write to buffer_size_kb can permanently fail, due to cpu_online_mask may changed between two for_each_online_buffer_cpu loops. The number of increasing and decreasing on cpu_buffer-

...

resize_disable

may be inconsistent, leading that the resize_disabled in some CPUs becoming none zero after ring_buffer_reset_online_cpus return.

This issue can be reproduced by "echo 0 > trace" while hotplugging cpu. After reproducing success, we can find out buffer_size_kb will not be functional anymore.

Prevent the two "loops" in this function from iterating through different online cpus by copying cpu_online_mask at the entry of the function.

The "Changes from" need to go below the '---', otherwise they are added to the git commit (we don't want it there).

Will remember this, won't happened next time :)

...

Changes from v1 to v3: Declare the cpumask variable statically rather than dynamically.

Changes from v2 to v3: Considering holding cpu_hotplug_lock too long because of the synchronize_rcu(), maybe it's better to prevent the issue by copying cpu_online_mask at the entry of the function as V1 does, instead of using cpus_read_lock().

Link: https://lore.kernel.org/lkml/20230408052226.25268-1-Tze-nan.Wu@mediatek.com/ Link: https://lore.kernel.org/oe-kbuild-all/202304082051.Dp50upfS-lkp@intel.com/ Link: https://lore.kernel.org/oe-kbuild-all/202304081615.eiaqpbV8-lkp@intel.com/

Cc: stable@vger.kernel.org Cc: npiggin@gmail.com Fixes: b23d7a5f4a07 ("ring-buffer: speed up buffer resets by avoiding synchronize_rcu for each CPU") Reported-by: kernel test robot lkp@intel.com Reviewed-by: Cheng-Jui Wang cheng-jui.wang@mediatek.com Signed-off-by: Tze-nan Wu Tze-nan.Wu@mediatek.com

---

This is where the "Changes from" go. And since this patch is not suppose to be a Cc. But since it's still good to have a link to it. You could do:

Changes from v2 to v3: https://lore.kernel.org/linux-trace-kernel/20230409024616.31099-1-Tze-nan.Wu... Considering holding cpu_hotplug_lock too long because of the synchronize_rcu(), maybe it's better to prevent the issue by copying cpu_online_mask at the entry of the function as V1 does, instead of using cpus_read_lock().

Where the previous version changes has the lore link to the previous patch, in case someone wants to look at it.

Sure, a link here is really helpful. Will follow this format in the future.

...

kernel/trace/ring_buffer.c | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c index 76a2d91eecad..dc758930dacb 100644 --- a/kernel/trace/ring_buffer.c +++ b/kernel/trace/ring_buffer.c @@ -288,9 +288,6 @@ EXPORT_SYMBOL_GPL(ring_buffer_event_data); #define for_each_buffer_cpu(buffer, cpu) \ for_each_cpu(cpu, buffer->cpumask)

-#define for_each_online_buffer_cpu(buffer, cpu) \

• for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
  

#define TS_SHIFT 27 #define TS_MASK ((1ULL << TS_SHIFT) - 1) #define TS_DELTA_TEST (~TS_MASK) @@ -5353,12 +5350,19 @@ EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu); void ring_buffer_reset_online_cpus(struct trace_buffer *buffer) { struct ring_buffer_per_cpu *cpu_buffer;

• cpumask_t reset_online_cpumask;
  

It's usually considered bad form to put a cpumask on the stack. As it can be 128 bytes for a machine with 1024 CPUs (and yes they do exist). Also, the mask size is set to NR_CPUS not the actual size, so you do not even need to have it that big.

Never thought about that until you told me, I will keep it in mind before declare a cpumask next time.

...

  int cpu;

• /*
  

•  * Record cpu_online_mask here to make sure we iterate
  

through the same

•  * online CPUs in the following two loops.
  

•  */
  

• cpumask_copy(&reset_online_cpumask, cpu_online_mask);
  

• /* prevent another thread from changing buffer sizes */
  mutex_lock(&buffer->mutex);
  
  

• for_each_online_buffer_cpu(buffer, cpu) {
  

• for_each_cpu_and(cpu, buffer->cpumask, &reset_online_cpumask)
  

{ cpu_buffer = buffer->buffers[cpu];

          atomic_inc(&cpu_buffer->resize_disabled);

Anyway, we don't need to modify any of the above, and just do the following instead of atomic_inc():

#define RESET_BIT (1 << 30)

            atomic_add(&cpu_buffer->resize_disabled, RESET_BIT);

...

@@ -5368,7 +5372,7 @@ void ring_buffer_reset_online_cpus(struct trace_buffer *buffer) /* Make sure all commits have finished */ synchronize_rcu();

• for_each_online_buffer_cpu(buffer, cpu) {
  

• for_each_cpu_and(cpu, buffer->cpumask, &reset_online_cpumask)
  

Maybe we should use for_each_buffer_cpu(buffer, cpu) here? since a CPU may also came offline during synchronize_rcu().

...

{ cpu_buffer = buffer->buffers[cpu];

Then here we can do:

            /*
             * If a CPU came online during the synchronize_rcu(),

then * ignore it. */ if (!atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT)) continue;

            atomic_sub(&cpu_buffer->resize_disabled, RESET_BIT);

As the resize_disabled only needs to be set to something to make it disabled.

-- Steve

Thanks for all your suggestions, learn a lot from here, really appriciate :). I will upload a v4 patch in new thread as soon as the new patch pass my test.

-- Tzenan

...

          reset_disabled_cpu_buffer(cpu_buffer);