Re: [PATCH 2/2] /proc/stat: Add sysctl parameter to control irq counts latency

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 01/07/2019 10:58 AM, Matthew Wilcox wrote:
> On Mon, Jan 07, 2019 at 10:12:58AM -0500, Waiman Long wrote:
>> Reading /proc/stat can be slow especially if there are many irqs and on
>> systems with many CPUs as summation of all the percpu counts for each
>> of the irqs is required. On some newer systems, there can be more than
>> 1000 irqs per socket.
>>
>> Applications that need to read /proc/stat many times per seconds will
>> easily hit a bottleneck. In reality, the irq counts are seldom looked
>> at. Even those applications that read them don't really need up-to-date
>> information. One way to reduce the performance impact of irq counts
>> computation is to do it less frequently.
>>
>> A new "fs/proc-stat-irqs-latency-ms" sysctl parameter is now added to
> No.  No, no, no, no, no.  No.
>
> Stop adding new sysctls for this kind of thing.  It's just a way to shift
> blame from us (who allegedly know what we're doing) to the sysadmin
> (who probably has better things to be doing than keeping up with the
> intricacies of development of every single piece of software running on
> their system).
>
> Let's figure out what this _should_ be.  As a strawman, I propose we
> update these stats once a second.  That's easy to document and understand.

I am fine with having a fixed value (like 1s) for reporting purpose. It
is just people may have many different opinions on what the right value
should be. That is why I opt for flexibility in the initial patch.

>
>> @@ -98,7 +105,48 @@ static u64 compute_stat_irqs_sum(void)
>>  static void show_stat_irqs(struct seq_file *p)
>>  {
>>  	int i;
>> +#ifdef CONFIG_PROC_SYSCTL
>> +	static char *irqs_buf;		   /* Buffer for irqs values */
>> +	static int buflen;
>> +	static unsigned long last_jiffies; /* Last buffer update jiffies */
>> +	static DEFINE_MUTEX(irqs_mutex);
>> +	unsigned int latency = proc_stat_irqs_latency_ms;
>> +
>> +	if (latency) {
>> +		char *ptr;
>> +
>> +		latency = _msecs_to_jiffies(latency);
>> +
>> +		mutex_lock(&irqs_mutex);
>> +		if (irqs_buf && time_before(jiffies, last_jiffies + latency))
>> +			goto print_out;
>> +
>> +		/*
>> +		 * Each irq value may require up to 11 bytes.
>> +		 */
>> +		if (!irqs_buf) {
>> +			irqs_buf = kmalloc(nr_irqs * 11 + 32,
>> +					   GFP_KERNEL | __GFP_ZERO);
> Why are you caching the _output_ of calling sprintf(), rather than caching the
> values of each interrupt?
>
It is just faster to dump the whole string buffer than redoing the
number formatting each time when the values don't change. I can cache
the individual sums instead if it is the preferred by most.

Cheers,
Longman





[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux