答复: [????] Re: [PATCH] nodemask: use nr_node_ids instead of MAX_NUMNODES in __nodemask_pr_numnodes()

["Li,Rongqing" <lirongqing@baidu.com>]

> > On Thu, 9 Jul 2026 15:17:39 +0800 lirongqing <lirongqing@baidu.com>
> > wrote:
> >
> > > From: Li RongQing <lirongqing@baidu.com>
> > >
> > > __nodemask_pr_numnodes() returns MAX_NUMNODES, a compile-time
> > > constant, as the width for '%*pb[l]' nodemask printing. On systems
> > > with few NUMA nodes this produces excessive zero-padding, e.g. with
> > > MAX_NUMNODES=1024 but only 2 nodes:
> > >
> > >   Mems_allowed: 00000000,00000000,...,00000003  (32 groups)
> >
> > Please let's completely describe the effects of this.
> > nodemask_pr_args() is used in several places so I believe that a
> > number of userspace-facing strings will now appear differently?  Let's
> > spell out which userspace interfaces are affected, and how.
> >
> 
> This affects every nodemask_pr_args() user.  Userspace-visible output
> includes node state sysfs files under /sys/devices/system/node/, NVMe
> multipath numa_nodes, memory tier nodelist, cpuset mems/effective_mems
> files, /proc/<pid>/status Mems_allowed*, mempolicy strings such as those
> shown in /proc/<pid>/numa_maps, and slub debugfs "nodes=" output.
> Kernel log messages that print nodemasks, such as cpuset warnings,
> allocation failure warnings and OOM diagnostics, are affected as well.
> 
> The semantic value is unchanged.  Only the printed width changes from
> MAX_NUMNODES bits to nr_node_ids bits, removing trailing zero words on
> systems built with a large MAX_NUMNODES but booted with a smaller
> possible-node range.  List outputs (%*pbl) remain unchanged for common
> contiguous node ranges such as "0-1"; bitmap outputs (%*pb), such as
> /proc/<pid>/status Mems_allowed, become shorter.
> 

Sorry, The upper reply is wrong.

After checking all call sites of nodemask_pr_args(), I confirm that the excessive zero-padding only affects the Mems_allowed: line in /proc/<pid>/status.

The reason is that nodemask_pr_args() is used with two different format specifiers:

%*pbl (list format, 9 out of 10 callers) ― prints nodemask as compact ranges like 0-1. The width from __nodemask_pr_numnodes() is effectively ignored by bitmap_list_string(), so these are unaffected.
%*pb (bitmap format, 1 caller) ― prints nodemask as a hex bitmap. The width directly controls the number of hex groups output. This is only used in cpuset_task_status_allowed() for the Mems_allowed: field in /proc/<pid>/status.
So the visible change is limited to grep Mems_allowed /proc/1/status (and equivalent for other PIDs). No other userspace-facing strings are affected. 

Thanks


[Li,Rongqing] 



> 
> > > Use nr_node_ids (the runtime highest-node-id + 1) instead, matching
> > > the behavior of cpumask_pr_args() which uses nr_cpu_ids:
> > >
> > >   Mems_allowed: 00000003
> > >
> > > Move the nr_node_ids declaration earlier in the file (guarded by
> > > __nodemask_pr_numnodes(), and remove the now-duplicate
> declarations.
> > >
> > > ...
> > >
> > > --- a/include/linux/nodemask.h
> > > +++ b/include/linux/nodemask.h
> > > @@ -95,6 +95,12 @@
> > >
> > >  extern nodemask_t _unused_nodemask_arg_;
> > >
> > > +#if MAX_NUMNODES > 1
> > > +extern unsigned int nr_node_ids;
> > > +#else
> > > +#define nr_node_ids		1U
> > > +#endif
> > > +
> > >  /**
> > >   * nodemask_pr_args - printf args to output a nodemask
> > >   * @maskp: nodemask to be printed
> > > @@ -105,7 +111,7 @@ extern nodemask_t _unused_nodemask_arg_;
> > >  				__nodemask_pr_bits(maskp)
> > >  static __always_inline unsigned int __nodemask_pr_numnodes(const
> > > nodemask_t *m)  {
> > > -	return m ? MAX_NUMNODES : 0;
> > > +	return m ? nr_node_ids : 0;
> > >  }
> > >  static __always_inline const unsigned long
> > > *__nodemask_pr_bits(const nodemask_t *m)  { @@ -438,7 +444,6 @@
> > > static __always_inline
> > unsigned
> > > int next_memory_node(int nid)
> > >  	return next_node(nid, node_states[N_MEMORY]);  }
> > >
> > > -extern unsigned int nr_node_ids;
> > >  extern unsigned int nr_online_nodes;
> >
> > Let's relocate nr_online_nodes also?  To keep them together.
> 
> 
> I will relocate it in V2
> 
> Thanks
> 
> [Li,Rongqing]
> 
> 
> 
> >
> > >  static __always_inline void node_set_online(int nid) @@ -480,7
> > > +485,6 @@ static __always_inline int num_node_state(enum node_states
> state)
> > >  #define first_memory_node	0
> > >  #define next_online_node(nid)	(MAX_NUMNODES)
> > >  #define next_memory_node(nid)	(MAX_NUMNODES)
> > > -#define nr_node_ids		1U
> > >  #define nr_online_nodes		1U
> > >
> > >  #define node_set_online(node)	   node_set_state((node), N_ONLINE)
> > > --
> > > 2.9.4