#10879 NORM Opportu: busybox rm not as quick as coreutils

[Zarro Boogs per Child]

#10879: busybox rm not as quick as coreutils
---------------------------------+------------------------------------------
           Reporter:  dsd        |       Owner:  dsd          
               Type:  defect     |      Status:  new          
           Priority:  normal     |   Milestone:  Opportunity  
          Component:  initramfs  |     Version:  not specified
         Resolution:             |    Keywords:               
        Next_action:  never set  |    Verified:  0            
Deployment_affected:             |   Blockedby:               
           Blocking:             |  
---------------------------------+------------------------------------------

Comment(by pgf):

 i was curious, so i did some timing and stracing on a server.  i see a
 similar performance ratio between /bin/rm and busybox rm as reported
 above.

 busybox rm seems to traverse the tree from the parent, using string
 concatenation to form pathnames.  this results in
 {{{
 unlink("s/contrib/doc/shpproj.txt")
 }}}

 coreutils rm seems to open subdirectories directly (using openat()), and
 performs the unlinks using "unlinkat()".  this avoids both string
 manipulation in the command, and repeated duplicate path parsing in the
 kernel.  i.e., the above unlink looks like this in coreutils:

 {{{
 unlinkat(3, "shpproj.txt", 0)
 }}}
 where fd 3 came from an openat() of the "doc" directory.

 (i did the timings with a git checkout of busybox 1.15.1.  the timings
 were the same as git HEAD,
 which seems to be >= 1.18.4.  the stracing was done with HEAD.)

-- 
Ticket URL: <http://dev.laptop.org/ticket/10879#comment:1>
One Laptop Per Child <http://laptop.org/>
OLPC bug tracking system