[PATCH] mm: simplify node/zone name printing

[PATCH] mm: simplify node/zone name printing

[Michal Hocko]

From: Michal Hocko <mhocko@suse.com>

show_node currently only prints Node id while it is always followed by
printing zone->name. As the node information is conditional to
CONFIG_NUMA we have to be careful to always terminate the previous
continuation line before printing the zone name. This is quite ugly
and easy to mess up. Let's rename show_node to show_zone_node and
make sure that it will always start at a new line. We can drop the ugly
printk(KERN_CONT "\n") from show_free_areas.

Signed-off-by: Michal Hocko <mhocko@suse.com>
---
Hi,
this has been sitting in my tree since oct and I completely forgot about
it. Does this look like a reasonable clean up to you?

 mm/page_alloc.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 3f2c9e535f7f..5324efa8b9d0 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -4120,10 +4120,12 @@ unsigned long nr_free_pagecache_pages(void)
 	return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
 }
 
-static inline void show_node(struct zone *zone)
+static inline void show_zone_node(struct zone *zone)
 {
 	if (IS_ENABLED(CONFIG_NUMA))
-		printk("Node %d ", zone_to_nid(zone));
+		printk("Node %d %s", zone_to_nid(zone), zone->name);
+	else
+		printk("%s: ", zone->name);
 }
 
 long si_mem_available(void)
@@ -4371,9 +4373,8 @@ void show_free_areas(unsigned int filter)
 		for_each_online_cpu(cpu)
 			free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
 
-		show_node(zone);
+		show_zone_node(zone);
 		printk(KERN_CONT
-			"%s"
 			" free:%lukB"
 			" min:%lukB"
 			" low:%lukB"
@@ -4396,7 +4397,6 @@ void show_free_areas(unsigned int filter)
 			" local_pcp:%ukB"
 			" free_cma:%lukB"
 			"\n",
-			zone->name,
 			K(zone_page_state(zone, NR_FREE_PAGES)),
 			K(min_wmark_pages(zone)),
 			K(low_wmark_pages(zone)),
@@ -4421,7 +4421,6 @@ void show_free_areas(unsigned int filter)
 		printk("lowmem_reserve[]:");
 		for (i = 0; i < MAX_NR_ZONES; i++)
 			printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
-		printk(KERN_CONT "\n");
 	}
 
 	for_each_populated_zone(zone) {
@@ -4431,8 +4430,7 @@ void show_free_areas(unsigned int filter)
 
 		if (skip_free_areas_node(filter, zone_to_nid(zone)))
 			continue;
-		show_node(zone);
-		printk(KERN_CONT "%s: ", zone->name);
+		show_zone_node(zone);
 
 		spin_lock_irqsave(&zone->lock, flags);
 		for (order = 0; order < MAX_ORDER; order++) {
-- 
2.10.2

Re: [PATCH] mm: simplify node/zone name printing

[Vlastimil Babka]

On 12/16/2016 01:32 PM, Michal Hocko wrote:
> From: Michal Hocko <mhocko@suse.com>
> 
> show_node currently only prints Node id while it is always followed by
> printing zone->name. As the node information is conditional to
> CONFIG_NUMA we have to be careful to always terminate the previous
> continuation line before printing the zone name. This is quite ugly
> and easy to mess up. Let's rename show_node to show_zone_node and
> make sure that it will always start at a new line. We can drop the ugly
> printk(KERN_CONT "\n") from show_free_areas.
> 
> Signed-off-by: Michal Hocko <mhocko@suse.com>

Acked-by: Vlastimil Babka <vbabka@suse.cz>

Just a question below... (CC printk experts)

> ---
> Hi,
> this has been sitting in my tree since oct and I completely forgot about
> it. Does this look like a reasonable clean up to you?

Yeah, even besides the removed line, which my question is about....

>  mm/page_alloc.c | 14 ++++++--------
>  1 file changed, 6 insertions(+), 8 deletions(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 3f2c9e535f7f..5324efa8b9d0 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -4120,10 +4120,12 @@ unsigned long nr_free_pagecache_pages(void)
>  	return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
>  }
>  
> -static inline void show_node(struct zone *zone)
> +static inline void show_zone_node(struct zone *zone)
>  {
>  	if (IS_ENABLED(CONFIG_NUMA))
> -		printk("Node %d ", zone_to_nid(zone));
> +		printk("Node %d %s", zone_to_nid(zone), zone->name);
> +	else
> +		printk("%s: ", zone->name);
>  }
>  
>  long si_mem_available(void)
> @@ -4371,9 +4373,8 @@ void show_free_areas(unsigned int filter)
>  		for_each_online_cpu(cpu)
>  			free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
>  
> -		show_node(zone);
> +		show_zone_node(zone);
>  		printk(KERN_CONT
> -			"%s"
>  			" free:%lukB"
>  			" min:%lukB"
>  			" low:%lukB"
> @@ -4396,7 +4397,6 @@ void show_free_areas(unsigned int filter)
>  			" local_pcp:%ukB"
>  			" free_cma:%lukB"
>  			"\n",
> -			zone->name,
>  			K(zone_page_state(zone, NR_FREE_PAGES)),
>  			K(min_wmark_pages(zone)),
>  			K(low_wmark_pages(zone)),
> @@ -4421,7 +4421,6 @@ void show_free_areas(unsigned int filter)
>  		printk("lowmem_reserve[]:");
>  		for (i = 0; i < MAX_NR_ZONES; i++)
>  			printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
> -		printk(KERN_CONT "\n");

So there's really no functional difference between terminating line
explicitly with "\n", and doing a followup printk() without KERN_CONT?
I agree that a KERN_CONT line just to print "\n" is ugly, just want to
be sure we are really safe without it, considering how KERN_CONT has
been recently changed etc.

>  	}
>  
>  	for_each_populated_zone(zone) {
> @@ -4431,8 +4430,7 @@ void show_free_areas(unsigned int filter)
>  
>  		if (skip_free_areas_node(filter, zone_to_nid(zone)))
>  			continue;
> -		show_node(zone);
> -		printk(KERN_CONT "%s: ", zone->name);
> +		show_zone_node(zone);
>  
>  		spin_lock_irqsave(&zone->lock, flags);
>  		for (order = 0; order < MAX_ORDER; order++) {
> 

Re: [PATCH] mm: simplify node/zone name printing

[Sergey Senozhatsky]

Hello,

On (12/19/16 08:00), Vlastimil Babka wrote:
[..]
> > @@ -4421,7 +4421,6 @@ void show_free_areas(unsigned int filter)
> >  		printk("lowmem_reserve[]:");
> >  		for (i = 0; i < MAX_NR_ZONES; i++)
> >  			printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
> > -		printk(KERN_CONT "\n");
> 
> So there's really no functional difference between terminating line
> explicitly with "\n", and doing a followup printk() without KERN_CONT?
> I agree that a KERN_CONT line just to print "\n" is ugly, just want to
> be sure we are really safe without it, considering how KERN_CONT has
> been recently changed etc.

depending on what will follow. if another KERN_CONT printk then
without this explicit '\n' printk may assume that we are still
in cont printing; if something else -- printk should flush.

kernel/printk/printk.c

	/*
	 * If an earlier line was buffered, and we're a continuation
	 * write from the same process, try to add it to the buffer.
	 */
	if (cont.len) {
		if (cont.owner == current && (lflags & LOG_CONT)) {
			if (cont_add(facility, level, lflags, text, text_len))
				return text_len;
		}
		/* Otherwise, make sure it's flushed */
		cont_flush();
	}


as far as I can tell, now for_each_populated_zone() iterations are
split by non-CONT printk() from show_zone_node(), which previously
has been   printk(KERN_CONT "%s: ", zone->name), so pr_cont(\n)
between iterations was important, but now that non-CONT printk()
should do the trick. it's _a bit_ hacky, though.

	-ss

> >  	}
> >  
> >  	for_each_populated_zone(zone) {
> > @@ -4431,8 +4430,7 @@ void show_free_areas(unsigned int filter)
> >  
> >  		if (skip_free_areas_node(filter, zone_to_nid(zone)))
> >  			continue;
> > -		show_node(zone);
> > -		printk(KERN_CONT "%s: ", zone->name);
> > +		show_zone_node(zone);
> >  
> >  		spin_lock_irqsave(&zone->lock, flags);

Re: [PATCH] mm: simplify node/zone name printing

[Michal Hocko]

On Mon 19-12-16 16:32:28, Sergey Senozhatsky wrote:
[...]
> as far as I can tell, now for_each_populated_zone() iterations are
> split by non-CONT printk() from show_zone_node(), which previously
> has been   printk(KERN_CONT "%s: ", zone->name), so pr_cont(\n)
> between iterations was important, but now that non-CONT printk()
> should do the trick. it's _a bit_ hacky, though.

Do you consider that more hacky than the original? At least for me,
starting with KERN_CONT and relying on an explicit \n sounds more error
prone than leaving the last pr_cont without \n and relying on the
implicit flushing.
-- 
Michal Hocko
SUSE Labs

Re: [PATCH] mm: simplify node/zone name printing

[Sergey Senozhatsky]

On (12/19/16 09:12), Michal Hocko wrote:
> On Mon 19-12-16 16:32:28, Sergey Senozhatsky wrote:
> [...]
> > as far as I can tell, now for_each_populated_zone() iterations are
> > split by non-CONT printk() from show_zone_node(), which previously
> > has been   printk(KERN_CONT "%s: ", zone->name), so pr_cont(\n)
> > between iterations was important, but now that non-CONT printk()
> > should do the trick. it's _a bit_ hacky, though.
> 
> Do you consider that more hacky than the original?

well, slightly. merely because (to me) implicit always relies on
internals, which can change; while explicit does not (ideally).
simply because of that.

but I don't have any problems with your patch.

	-ss

Re: [PATCH] mm: simplify node/zone name printing

[Petr Mladek]

On Mon 2016-12-19 08:00:47, Vlastimil Babka wrote:
> On 12/16/2016 01:32 PM, Michal Hocko wrote:
> > From: Michal Hocko <mhocko@suse.com>
> > 
> > show_node currently only prints Node id while it is always followed by
> > printing zone->name. As the node information is conditional to
> > CONFIG_NUMA we have to be careful to always terminate the previous
> > continuation line before printing the zone name. This is quite ugly
> > and easy to mess up. Let's rename show_node to show_zone_node and
> > make sure that it will always start at a new line. We can drop the ugly
> > printk(KERN_CONT "\n") from show_free_areas.
> > 
> > Signed-off-by: Michal Hocko <mhocko@suse.com>
> 
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 3f2c9e535f7f..5324efa8b9d0 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -4120,10 +4120,12 @@ unsigned long nr_free_pagecache_pages(void)
> >  	return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
> >  }
> >  
> > -static inline void show_node(struct zone *zone)
> > +static inline void show_zone_node(struct zone *zone)
> >  {
> >  	if (IS_ENABLED(CONFIG_NUMA))
> > -		printk("Node %d ", zone_to_nid(zone));
> > +		printk("Node %d %s", zone_to_nid(zone), zone->name);
> > +	else
> > +		printk("%s: ", zone->name);
> >  }
> >  
> >  long si_mem_available(void)
> > @@ -4371,9 +4373,8 @@ void show_free_areas(unsigned int filter)
> >  		for_each_online_cpu(cpu)
> >  			free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
> >  
> > -		show_node(zone);
> > +		show_zone_node(zone);
> >  		printk(KERN_CONT
> > -			"%s"

The new code will printk "%s: " when called with disabled CONFIG_NUMA.
Is the added ": " OK?

Best Regards,
Petr

Re: [PATCH] mm: simplify node/zone name printing

[Petr Mladek]

On Mon 2016-12-19 09:12:10, Michal Hocko wrote:
> On Mon 19-12-16 16:32:28, Sergey Senozhatsky wrote:
> [...]
> > as far as I can tell, now for_each_populated_zone() iterations are
> > split by non-CONT printk() from show_zone_node(), which previously
> > has been   printk(KERN_CONT "%s: ", zone->name), so pr_cont(\n)
> > between iterations was important, but now that non-CONT printk()
> > should do the trick. it's _a bit_ hacky, though.
> 
> Do you consider that more hacky than the original? At least for me,
> starting with KERN_CONT and relying on an explicit \n sounds more error
> prone than leaving the last pr_cont without \n and relying on the
> implicit flushing.

The missing '\n' will cause the string will not be flushed
until another printk happens. It is not a problem here because
other printk follows. But it might take a while in general.

There was a commit[1] that flushed the cont lines when the log
buffer was read via /dev/kmsg or syslog. Also there was a patch[2]
that flushed cont lines using a timer. But the commit caused problems
and was reverted[3]. Also the patch needs more testing. So, it might
take a while until flushing partial cont lines is "guaranteed".

I would personally prefer to keep that pr_cont("\n") call when
it was already there. It makes it clear that the line is over.
It does not rely on any further printk's or fallbacks.

[1] https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=bfd8d3f23b51018388be0411ccbc2d56277fe294
[2]
https://lkml.kernel.org/r/CA+55aFwKYnrMJr_vSE+GfDGszeUGyd=CPUD15-zZ8yWQW61GBA@mail.gmail.com
[3] https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=f5c9f9c72395c3291c2e35c905dedae2b98475a4

Best Regards,
Petr

Re: [PATCH] mm: simplify node/zone name printing

[Michal Hocko]

On Mon 19-12-16 11:05:49, Petr Mladek wrote:
> On Mon 2016-12-19 08:00:47, Vlastimil Babka wrote:
> > On 12/16/2016 01:32 PM, Michal Hocko wrote:
> > > From: Michal Hocko <mhocko@suse.com>
> > > 
> > > show_node currently only prints Node id while it is always followed by
> > > printing zone->name. As the node information is conditional to
> > > CONFIG_NUMA we have to be careful to always terminate the previous
> > > continuation line before printing the zone name. This is quite ugly
> > > and easy to mess up. Let's rename show_node to show_zone_node and
> > > make sure that it will always start at a new line. We can drop the ugly
> > > printk(KERN_CONT "\n") from show_free_areas.
> > > 
> > > Signed-off-by: Michal Hocko <mhocko@suse.com>
> > 
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 3f2c9e535f7f..5324efa8b9d0 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -4120,10 +4120,12 @@ unsigned long nr_free_pagecache_pages(void)
> > >  	return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
> > >  }
> > >  
> > > -static inline void show_node(struct zone *zone)
> > > +static inline void show_zone_node(struct zone *zone)
> > >  {
> > >  	if (IS_ENABLED(CONFIG_NUMA))
> > > -		printk("Node %d ", zone_to_nid(zone));
> > > +		printk("Node %d %s", zone_to_nid(zone), zone->name);
> > > +	else
> > > +		printk("%s: ", zone->name);
> > >  }
> > >  
> > >  long si_mem_available(void)
> > > @@ -4371,9 +4373,8 @@ void show_free_areas(unsigned int filter)
> > >  		for_each_online_cpu(cpu)
> > >  			free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
> > >  
> > > -		show_node(zone);
> > > +		show_zone_node(zone);
> > >  		printk(KERN_CONT
> > > -			"%s"
> 
> The new code will printk "%s: " when called with disabled CONFIG_NUMA.
> Is the added ": " OK?

no, that was not intentional. Will drop it.

-- 
Michal Hocko
SUSE Labs

Re: [PATCH] mm: simplify node/zone name printing

[Michal Hocko]

On Mon 19-12-16 11:27:59, Petr Mladek wrote:
> On Mon 2016-12-19 09:12:10, Michal Hocko wrote:
> > On Mon 19-12-16 16:32:28, Sergey Senozhatsky wrote:
> > [...]
> > > as far as I can tell, now for_each_populated_zone() iterations are
> > > split by non-CONT printk() from show_zone_node(), which previously
> > > has been   printk(KERN_CONT "%s: ", zone->name), so pr_cont(\n)
> > > between iterations was important, but now that non-CONT printk()
> > > should do the trick. it's _a bit_ hacky, though.
> > 
> > Do you consider that more hacky than the original? At least for me,
> > starting with KERN_CONT and relying on an explicit \n sounds more error
> > prone than leaving the last pr_cont without \n and relying on the
> > implicit flushing.
> 
> The missing '\n' will cause the string will not be flushed
> until another printk happens. It is not a problem here because
> other printk follows. But it might take a while in general.
> 
> There was a commit[1] that flushed the cont lines when the log
> buffer was read via /dev/kmsg or syslog. Also there was a patch[2]
> that flushed cont lines using a timer. But the commit caused problems
> and was reverted[3]. Also the patch needs more testing. So, it might
> take a while until flushing partial cont lines is "guaranteed".

OK, fair enough. If the flushing partial cont lines is not guaranteed
then this cleanup makes less sense. I was under impression this has been
solved but as per your list of commits we are not there yet.

So let's just drop this patch and I will try to remember to double check
later.

Thanks!

-- 
Michal Hocko
SUSE Labs