On Wed, Nov 29, 2017 at 10:12:09PM -0500, Steven Rostedt wrote:
> On Wed, 29 Nov 2017 12:42:45 +0800
> changbin...@intel.com wrote:
>
> > From: Changbin Du <changbin...@intel.com>
> >
> > The default NR_CPUS can be very large, but actual possible nr_cpu_ids
> > usually is very small. For my x86 distribution, the NR_CPUS is 8192 and
> > nr_cpu_ids is 4. About 2 pages are wasted.
> >
> > Most machines don't have so many CPUs, so define a array with NR_CPUS
> > just wastes memory. So let's allocate the buffer dynamically when need.
> >
> > The exact buffer size should be:
> > DIV_ROUND_UP(nr_cpu_ids, 4) + nr_cpu_ids/32 + 2;
> >
> > Example output:
> > ff,ffffffff
>
> Um, what if there's more than 64 CPUs, where I have booted several
> boxes that have more. There's going to be more than 1 comma.
>
The commas are calculated by formula. (DIV_ROUND_UP(nr_cpu_ids, 4))
> >
> > With this change, the mutext tracing_cpumask_update_lock also can be
> > removed now, which was used to protect mask_str.
> >
> > Signed-off-by: Changbin Du <changbin...@intel.com>
> > Cc: Steven Rostedt <rost...@goodmis.org>
> >
> > ---
> > v3:
> > - remove tracing_cpumask_update_lock which was used to protect mask_str.
> > (Rostedt)
> > v2:
> > - remove 'static' declaration.
> > - fix buffer size.
> > ---
> > kernel/trace/trace.c | 29 +++++++++--------------------
> > 1 file changed, 9 insertions(+), 20 deletions(-)
> >
> > diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> > index 73e67b6..6750d05 100644
> > --- a/kernel/trace/trace.c
> > +++ b/kernel/trace/trace.c
> > @@ -4178,37 +4178,30 @@ static const struct file_operations
> > show_traces_fops = {
> > .llseek = seq_lseek,
> > };
> >
> > -/*
> > - * The tracer itself will not take this lock, but still we want
> > - * to provide a consistent cpumask to user-space:
> > - */
> > -static DEFINE_MUTEX(tracing_cpumask_update_lock);
> > -
> > -/*
> > - * Temporary storage for the character representation of the
> > - * CPU bitmask (and one more byte for the newline):
> > - */
> > -static char mask_str[NR_CPUS + 1];
> > -
> > static ssize_t
> > tracing_cpumask_read(struct file *filp, char __user *ubuf,
> > size_t count, loff_t *ppos)
> > {
> > struct trace_array *tr = file_inode(filp)->i_private;
> > + char *mask_str;
> > int len;
> >
> > - mutex_lock(&tracing_cpumask_update_lock);
> > + /* Bitmap, ',' and two more bytes for the newline and '\0'. */
> > + len = DIV_ROUND_UP(nr_cpu_ids, 4) + nr_cpu_ids/32 + 2;
>
> This is broken. Instead do:
>
> len = snprintf(NULL, 0, "%*pb\n",
> cpumask_pr_args(tr->tracing_cpumask)) + 1;
>
> mask_str = kmalloc(len, GFP_KERNEL);
> [..]
> len = snprintf(mask_str, len, "%*pb\n",
> cpumask_pr_args(tr->tracing_cpumask));
>
> -- Steve
>
hmm. I never know that snprintf has such usage. This is much better than
calculating it by a formula.
> > + mask_str = kmalloc(len, GFP_KERNEL);
> > + if (!mask_str)
> > + return -ENOMEM;
> >
> > - len = snprintf(mask_str, count, "%*pb\n",
> > + len = snprintf(mask_str, len, "%*pb\n",
> > cpumask_pr_args(tr->tracing_cpumask));
> > if (len >= count) {
> > count = -EINVAL;
> > goto out_err;
> > }
> > - count = simple_read_from_buffer(ubuf, count, ppos, mask_str, NR_CPUS+1);
> > + count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len);
> >
> > out_err:
> > - mutex_unlock(&tracing_cpumask_update_lock);
> > + kfree(mask_str);
> >
> > return count;
> > }
> > @@ -4228,8 +4221,6 @@ tracing_cpumask_write(struct file *filp, const char
> > __user *ubuf,
> > if (err)
> > goto err_unlock;
> >
> > - mutex_lock(&tracing_cpumask_update_lock);
> > -
> > local_irq_disable();
> > arch_spin_lock(&tr->max_lock);
> > for_each_tracing_cpu(cpu) {
> > @@ -4252,8 +4243,6 @@ tracing_cpumask_write(struct file *filp, const char
> > __user *ubuf,
> > local_irq_enable();
> >
> > cpumask_copy(tr->tracing_cpumask, tracing_cpumask_new);
> > -
> > - mutex_unlock(&tracing_cpumask_update_lock);
> > free_cpumask_var(tracing_cpumask_new);
> >
> > return count;
>
--
Thanks,
Changbin Du
