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@xxxxxxxxxxxx> 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@xxxxxxxxxxxx/ > Link: https://lore.kernel.org/oe-kbuild-all/202304082051.Dp50upfS-lkp@xxxxxxxxx/ > Link: https://lore.kernel.org/oe-kbuild-all/202304081615.eiaqpbV8-lkp@xxxxxxxxx/ > > Cc: stable@xxxxxxxxxxxxxxx > Cc: npiggin@xxxxxxxxx > Fixes: b23d7a5f4a07 ("ring-buffer: speed up buffer resets by avoiding synchronize_rcu for each CPU") > Reported-by: kernel test robot <lkp@xxxxxxxxx> > Reviewed-by: Cheng-Jui Wang <cheng-jui.wang@xxxxxxxxxxxx> > Signed-off-by: Tze-nan Wu <Tze-nan.Wu@xxxxxxxxxxxx> > --- 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@xxxxxxxxxxxx/ 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);
