Embedded Linux Evolution | Turing Techtalk

Embedded Linux Evolution | Turing Techtalk

• 1.
  The  Evolu*on  of   Embedded  Linux   Patanjali  Somayaji   (patanjalisomayaji@gmail.com)  
• 2.
  Embedded  Linux   TiVo   Roku   Leapster  Explorer   Samsung  Smart  TV  Robonaut  2  
• 3.
  The  Early  Days   •  Tradi*onally,  Embedded  Systems  used  “roll-­‐your-­‐own”,   “homegrown”  or  “in-­‐house”  RTOSes   •  1980:  Arrival  of  the  first  general  purpose  commercial  embedded  OS   –  VRTX  (from  Hunter  and  Ready)     –  Used  in  the  Hubble  space  telescope  and  soWware-­‐controlled  radio  in   early  Motorola  Mobile  phones)     •  Mid-­‐1980s:  Increased  adop*on  of  commercial  RTOSes   –  VxWorks  from  WindRiver  became  the  market  leader   –  Other  solu*ons  –  QNX,  Greenhills  Integrity   –  Embedded  developer  value  add  was  in  the  App     –  Time-­‐to-­‐market  increasingly  important   •  Linux  v0.1  in  1991,  by  1995  supported  Alpha,  i386,  MIPS  and  Sparc   architectures;  and  by  1996  had  support  for  m68K  and  PPC  
• 4.
  The  Early  Days   •  1995:  Busybox  ini*al  release  to  enable  a  small  Linux  image  to  fit  on  a   1.44  MB  floppy   –  Started  by  Bruce  Perens  to  use  as  a  Debian  bootable  install  and  rescue   disk   •  1997:  Linux  Router  Project  –  first  embedded  Linux  project   •  1999:  Commercial  RTOS  industry  was  USD  1  Billion   •  1999:  The  first  year  you  could  buy  a  commercial  off  the  shelf   Embedded  Linux  Distribu*on  –  Montavista’s  Red  Hat  Linux  based   “Hard  Hat  Linux”   •  1999:  Linux  based  TiVo  DVR  announced  (PPC  403,  16  MB  RAM,  13  GB   HDD)   •  2000:  Embedded  Linux  Consor*um  (ELC)  –  now  transi*oned  to  the   Linux  Founda*on  star*ng  2005   •  CE  Linux  Forum  (CELF)  –  now  CE  working  group  of  the  Linux   Foumda*on  
• 5.
  Why  did  Embedded  Linux  grow  ?   •  “given  enough  eyeballs,  all  bugs  are  shallow”  -­‐  ESR   –  So  reliability,  which  a  concern  in  the  beginning,  faded  away   •  Linux  supported  MMUs  providing  memory  par**oning  and  virtual   memory   –  App  crashes  do  not  crash  the  system   –  Became  important  as  embedded  systems  grew  in  complexity   –  MMUs  became  common  in  low-­‐cost  32-­‐bit  embedded  systems   –  uCLinux  developed  for  MMU-­‐less  high  volume,  cost-­‐sensi*ve  markets   •  Royalty  free   –  No  license  free  per  unit  deployed,  unlike  commercial  RTOSes   –  Commercial  linux  is  usually  sold  on  a  developer-­‐seat  license  basis   –  Important  for  large  scale  embedded  deployment  which  is  cost  sensi*ve   •  No  dependency  on  a  single  OS  vendor  for  a  cri*cal  component  of   embedded  applica*on  
• 6.
  Why  did  Embedded  Linux  grow  ?   •  Networking  was  usually  a  paid-­‐for  extra  on  commercial  RTOSes.   Linux  was  built  with  networking  support  from  the  ground  up.   •  Open  source  alterna*ve  –  Net  BSD   –  NetBSD  was  mature,  open  source  and  gemng  established  at  the  same   *me   –  Linux  however  was  based  on  GPLv2  which  triggered  a  lot  of  give-­‐back   –  More  drivers  and  architectures  started  to  get  support,  and  there  was  a   massive  Network  effect   •  ARM  processors  were  heavily  adopted,  and  Linux  was  ported  to   various  SoCs  and  development  planorms   •  Linux  development  in  other  areas  e.g.  Servers  were  leveraged  in   Embedded  systems  e.g.  Networking,  CPU  Hotplug  
• 7.
  Embedded  Linux  -­‐  Networking   •  The  Linux  Router  Project  (LRP)  –  the   first  embedded  Linux  project  ?   •  Pioneering  features   –  Small  base  OS  footprint   –  A  simplified  packaging  system   –  Menu  based  system  and  package   configura*on   –  Strict  separa*on  of  vola*le,  non-­‐ vola*le,  Read  Only,  and  Read/Write   areas  of  the  root  hierarchy   –  Unpacked  and  run  from  ramdisk  or   run  directly  from  flash   –  A  system  to  commit  configura*on   changes  to  a  non-­‐vola*le  medium   (Disk/Flash)   •  Spawned  a  deluge  of  Embedded  Linux   Networking  Devices   Cisco  Linksys  WRT54GL  Wireless-­‐G   Broadband  Linux  Router  
• 8.
  “Every  age  had  its  own  kind  of  war,  its   own  limi*ng  condi*ons,  and  its  own   peculiar  preconcep*ons”  –  Carl  von   Clausewitz,  German-­‐Prussian  Soldier  &   Military  Strategist  
• 9.
  Real-­‐*me  Linux   • Different  approaches   –  TimeSys  introduces  proprietary  Linux/RT  1.0       –  FSM  Labs  RTLinux  (commercially  licensed)     •  Real-­‐*me  kernel  running  Linux  as  a  low-­‐priority  thread   •  Posix  API  to  run  real-­‐*me  threads  from  User  Space   •  Under  15  micro-­‐second  response  (Year  2000)   •  Acquired  by  WindRiver  in  2007   –  MontaVista  worked  on  kernel  pre-­‐emp*on   •  Meanwhile  the  Linux  Kernel  developers  were   working  on  their  solu*on  
• 10.
  Real-­‐*me  Linux   • 2000  –  Ingo  Molnar  and  Andrew  Morton  start  work  on  voluntary   pre-­‐empt  “CONFIG_PREEMPT_VOLUNTARY”   –  Reduce  worst-­‐case  interrupt  latencies   –  Break  lengthy  kernel  cycles  into  smaller  chunks  and  add  checks  to   common  causes  of  long  latencies   –  Rescheduling  if  an  interrupt  handler  has  set  the  “need_resched”  flag   –  Eventually  merged  in  Linux  Kernel  2.6   •  2001  –  Robert  Love  began  work  on  the  kernel  pre-­‐emp*on   “CONFIG_PREEMPT”   –  Ini*ally  based  on  MontaVista  patch       –  Target  low-­‐latency  apps  and  smooth  user  interac*on   –  Makes  kernel  pre-­‐emp*ble  like  user-­‐space,  allowing  higher  priority   tasks  to  run  -­‐  all  kernel  code  outside  of  spinlock-­‐protected  regions  and   interrupt  handlers  eligible  for  pre-­‐emp*on   –  Adopted  by  MontaVista  in  MVL  4.0  and  merged  with  Linux  Kernel  2.6  
• 11.
  Controlling  a  laser  with  Linux  is  crazy,  but  everyone  in   this  room  is  crazy  in  his  own  way.  So  if  you  want  to  use   Linux  to  control  an  industrial  welding  laser,  I  have  no   problem  with  your  using  PREEMPT_RT.”  -­‐  Linus  Torvalds  
• 12.
  Real-­‐*me  Linux   • 2005  –  Ingo  Molnar  and  others  began  work  on   CONFIG_PREEMPT_RT   –  Convert  Linux  Kernel  into  a  fully  pre-­‐emp*ble  kernel   •  In-­‐kernel  locking  primi*ves  made  pre-­‐emp*ble   •  Cri*cal  sec*ons  made  pre-­‐emp*ble  (crea*on  of  non-­‐pre-­‐ emp*ble  sec*ons  s*ll  possible)   •  Implemented  priority  inheritance  for  in-­‐kernel  spinlocks  and   semaphores   •  Converted  interrupt  handlers  to  pre-­‐emp*ble  kernel  threads   •  Converted  Linux  *mer  API  into  separate  infrastructures  for   high  resolu*on  *mers  and  *meouts     •  10  to  30  micro-­‐second  scheduler  and  interrupt  latencies    
• 13.
  Real-­‐*me  Linux   • Patches  for  Linux  2.6  Kernel  suppor*ng  Real-­‐*me   –  Preemp*ble  Interrupt  Handlers  in  Thread  Context   –  IRQ-­‐Disable  Virtualiza*on  for  Drivers   •  IRQ  threads  disabled  without  masking  hardware   –  Integrated  Kernel  Mutex  with  Priority  Inheritance  (PI)   •  Protec*on  of  Kernel  Cri*cal  Sec*ons  by  Preemp*ble  PI  Mutex   –  PI  Mutex  Subs*tuted  for  Non-­‐Preemp*ble  Kernel  (SMP)  Locks   •  Big  Kernel  Lock  (BKL)  converted  to  PI  Mutex   •  Spin-­‐Locks  converted  to  PI  Mutex   •  Read-­‐Write  Locks  converted  to  PI  Mutex   •  RCU  Preemp*on  Enhancements   –  High  Resolu*on  Timers   –  User-­‐Space  Mutex   •  Robustness/Dead-­‐owner/Priority  Queueing    
• 14.
  ARM  is  Everywhere   •  Smartphones   •  Inexpensive  evalua*on   boards   –  BeagleBoards   –  BeagleBones   –  Raspberry  Pi   •  Now  also  coming  for   Servers   –  AMD  to  release  its  first   “Seaule”  ARM  Server   Chip  in  2014   Source:  hup://arstechnica.com/business/2012/08/from-­‐ altair-­‐to-­‐ipad-­‐35-­‐years-­‐of-­‐personal-­‐computer-­‐market-­‐share/  
• 15.
  ARM  Linux   • Started  out  in  1994  as  a  port  of  the  1.0.x  kernel  to  the   Acorn  5000   •  1999  Linux  2.2.0  –  ARM  architecture  added  to  mainline   •  Variety  of  architectures,  Kernel  Patches  and  drivers     –  Hardware  varies  from  SoC  to  SoC   –  Development  planorm  divergence   •  “ARM  Wrestling”   –  Grew  to  be  a  massive  collec*on  of  “subplanorms”   –  Managed  independently  by  different  developers     –  “ARM  support  is  a  labyrinthine  maze  of  wheels  within   reinvented  wheels”   •  Linaro  formed  in  2010  to  drive  common  development  and   integra*on  
• 16.
  “I  think  the  more  serious  long-­‐term  issue  we  have  in  the   kernel  is  the  wild  and  crazy  embedded  planorm  code   (and  mostly  ARM  -­‐  not  because  ARM  is  in  any  way   fundamentally  crazier,  but  because  ARM  is  clearly  the   most  successful  embedded  planorm  by  far).  The   embedded  world  has  always  tended  to  eschew   standardized  planorms:  they've  been  resource   constrained  etc,  so  they've  done  very  tailored  chip  and   board  solu*ons,  and  felt  that  they  couldn't  afford  a  lot  of   planorm  abstrac*on.”  -­‐  Linus  Torvalds  –  May  2011  
• 17.
  Mobile  Linux  …   2001/2002:     Sharp  Zaurus   SL-­‐5500  running   OpenZaurus  and   OPIE   SA-­‐1110  CPU  @   206  MHz,  64  MiB   RAM,  16  MiB   NOR     2003:   Motorola  A760   ARM  925  chip  (an   Intel  PXA270)  at   316  MHz,  48  MiB   flash  storage,   QVGA  (240  x  320)   touch  screen     2008:   HTC/T-­‐Mobile  G1     Qualcomm  MSM7210A   (ARM11)  processor  running  @   528  MHz,  256  MiB  NAND  flash   storage,  192  MiB  RAM,  3.2-­‐inch   touchscreen  (320  x  480).   Launched  with  Android  1.0   based  on  modified  Linux  2.6.25     2005:   Nokia  N770  Tablet     TI  OMAP  1710  @  252  MHz   64  MB  RAM,  128  MB  Flash   4.1  inch  touchscreen  (800  x   480).  Launched  on  Maemo/ Linux    
• 18.
  …  and  Mobile  Linux  Consor*ums  
• 19.
  Brief  Interlude:  Embedded  Filesystems   •  Ini*al  configura*ons   –  Expensive  NOR  Flash  for  Cri*cal  storage   –  Cheaper  and  denser  NAND  Flash  for  larger  capaci*es   –  NAND  flash  used  in  consumer  devices   •  Flash  usage  needed  “wear-­‐leveling”  to  spread  the  block  erases  evenly   •  Ini*al  Linux  support  for  read-­‐only  FS  (no  need  for  wear-­‐levelling)   –  CramFS  &  SquishFS   •  Since  Flash  connected  as  raw  memory  devices,  Linux  needed  to  handle  intelligently   for  RW  via  the  Linux  MTD  Interface  (Memory  Technology  Device)   –  JFFS2,  YAFFS2,  LogFS,  UBIFS   •  Smartphones  evolved  to  using  eMMC     –  Large  volume  of  NAND  storage  with  a  controller   –  Connect  as  block  devices   –  Use  flash  vendor’s  flash  wear-­‐leveling  algorithms     •  2012:  F2FS     –  SoWware  layer  to  handle  “smartness”  outside  the  Flash  Transla*on  Layer  (FTL)   –  Targeted  at  SSDs,  eMMC,  SD  Cards  and  other  Flash  storage  with  FTL     –  Flash  wear-­‐leveling  in  controller  
• 20.
  Brief  Interlude:  Power  Management   •  Embedded  Linux  –  started  off  with  Routers,  etc.   “plugged  in”  devices  so  Power  Management  was  not  a   priority   •  PM  focused  mostly  on  Desktop/Laptop/Server  support     –  APM  (1992  Intel  &  MicrosoW)   •  PM  in  BIOS,  BIOS  HW  specific     •  Reduce  CPU,  turn-­‐off  Display,  disable  Harddisk  aWer  *me-­‐out   –  ACPI  (1996  v1.0  open  standard  industry  specifica*on)   •  OS  controls  PM,  BIOS  provides  API   •  Power  Management  cri*cal  for  Mobile  Phones  –   supported  on  ARM  SoCs  with  the  usual  fragmenta*on   •  Integra*ng  Power  Management  in  drivers  and  kernel  is   a  major  ac*vity  for  the  Mobile  OEM  
• 21.
  Brief  Interlude:  Power  Management   •  Power  Management  mechanics   –  System  Suspend  –  User  forces  system  to  sleep   –  CPU  idle  –  Idle  thread  triggers  sleep  states   –  CPU  freq  –  CPU  frequency  scaling   –  Run*me  PM  –  Device  Power  Management   –  CPU  Hotplug  –  Remove  a  CPU  from  the  running  system  (ini*ally  used  for   Server  SMP  Linux  became  important  for  mul*-­‐core  ARM)   •  Due  to  a  lack  of  a  standard  like  ACPI,  ARM  SoCs  support  is  fragmented   –  ARM  SoCs  expose  a  lot  of  HW  knobs  to  SoWware,  directly  controlled  by  the  OS   drivers   –  Each  SoC  vendor  exposes  a  superset  of  standard  ARM  power  states  for  fine-­‐ grained  control,  increasing  complexity  of  SoC  enablement  code  and  peripheral   drivers   –  Different  approaches  to  OS  Power  Management  frameworks  due  to  a  lack  of   design  pauerns,  lack  of  common  infrastructure  and  intrinsic  HW  differences   –  Linaro  approach  to  integrate  and  unify  
• 22.
  hups://wiki.linaro.org/WorkingGroups/ PowerManagement/Doc/Architecture  
• 23.
  Android  “wakelock”   • Default  state  is  to  suspend   •  Kernel  implementa*on  and  Android  user-­‐space  APIs   –  When  a  wakelock  of  type  “WAKE_LOCK_IDLE”  is  in  effect,   system  does  not  enter  Idle  (low  power)   –  When  a  wakelock  of  type  “WAKE_LOCK_SUSPEND”  is  in   effect,  system  will  not  suspend     •  Highly  controversial  feature,  mul*ple  patch   submissions  and  mailing  list  discussions  over  many   years   •  Compa*ble  solu*on  being  mainlined  in  later  Kernel  3.x   series  
• 24.
  Is  Android  the  new  “Embedded   Linux”  ?    
• 25.
  Android   “The  interes*ng  thing  about  Android’s  design  is  how  liule  we   modified  the  kernel.  Most  embedded  systems  on  which  I  have   worked  have  made  dras*c  changes  to  the  kernel,  only  to  leave   user-­‐space  alone  —  for  example,  a  heavily-­‐modified  “real*me”   kernel  but  X11  for  a  GUI.     Android  is  the  opposite:  Only  minimal  changes  to  the  kernel,   but  a  user-­‐space  wholly  unlike  that  of  any  other  Unix  system.   In  fact,  Android’s  user-­‐space  is  so  different  from  stock  Linux,   you  can  easily  say  that  Android  is  not  in  any  way  a  Linux   system,  except  for  the  kernel.”     -­‐  Robert  Love,  Google  Engineer  and  Kernel  Contributor   hup://www.forbes.com/sites/quora/2013/05/13/what-­‐are-­‐the-­‐major-­‐changes-­‐that-­‐android-­‐made-­‐to-­‐the-­‐linux-­‐kernel/  
• 26.
  Android   •  ashmem  (Android  Shared  Memory),  a  file-­‐based  shared  memory  system   •  Binder,  an  inter-­‐process  communica*on  (IPC)  and  remote  procedure  call   (RPC)  system   •  logger,  a  high-­‐speed  in-­‐kernel  logging  mechanism  op*mized  for  writes   •  Paranoid  Networking,  a  mechanism  to  restrict  network  I/O  to  certain   processes   •  pmem  (Physical  Memory),  a  driver  for  mapping  large  chunks  of  physical   memory  into  user-­‐space   •  Viking  Killer,  a  replacement  OOM  killer  that  implements  Android’s  “kill   least  recently  used  process”  logic  under  low  memory  condi*ons   •  wakelocks,  Android’s  unique  power  management  solu*on,  in  which  the   default  state  of  the  device  is  sleep  and  explicit  ac*on  is  required  (via  a   wakelock)  to  prevent  that   •  The  usual  assortment  of  drivers,  ARM  architecture  ports,  and  other   associated  low-­‐level  code  necessary  to  support  Android  on  any  given   device   hup://www.forbes.com/sites/quora/2013/05/13/what-­‐are-­‐the-­‐major-­‐changes-­‐that-­‐android-­‐made-­‐to-­‐the-­‐linux-­‐kernel/  
• 27.
  Android   •  Android  is  the  most  popular  Embedded  Linux   “distro”  today   •  From  hos*lity  to  gradual  adop*on   – E.g.  mainlining  of  selec*ve  Android  features  in   Linux  Kernel     – Ubuntu  “MIR”  display  server  across  Desktop/ Tablet/Phones  leverages  Android  display  drivers   and  architecture   – We  all  love  “adb”  
• 29.
  Looking  ahead  …  
• 30.
  Linaro   •  Not-­‐for-­‐profit  engineering   organiza*on  consolida*ng  and   op*mizing  open  source  Linux  soWware   and  tools  for  the  ARM  architecture   •  Focus  on  generic  ARM  technology   common  to  all  ARM  SoC  vendors   •  Focus  on  areas  that  interact  directly   with  silicon  like  Mul*media,  Graphics,   Power  Management,  Kernel  and  Boot   •  Maintenance,  up-­‐streaming  and   implementa*on  of  standards   •  Established  in  2010  by  ARM,  Freescale,   IBM,  Samsung,  ST-­‐Ericsson  and  Texas   Instruments  
• 31.
  Yocto  Project   • Not  an  embedded  Linux   distribu*on  –  Yocto  provides   tools  to  build  a  custom   embedded  Linux  distribu*on   •  Based  on  Poky,  OpenEmbedded   •  Provides  Linux  Kernel,  System   Libraries,  UI  Framework     •  Op*onal  Sato  user  interface   based  on  GNOME  Mobile  Stack    
• 32.
  Yocto  Project  
• 33.
  Genivi   •  Industry  non-­‐profit  alliance  to   drive  adop*on  of  open   source  In-­‐vehicle   Infotainment  development   planorm   •  Launched  in  Mar  2009   •  Linux-­‐based  kernel,  core   services,  middleware  and   open  applica*on  interfaces   •  Genivi  compliant  –  Tizen,   MontaVista  IVI,  Ubuntu   Remix  IVI,  WindRiver  IVI  
• 34.
  Tizen   •  History  –  builds  on  SLP  from   LiMo,  formed  when  Intel  leW   Meego  to  join  Tizen   •  Open-­‐source,  standards   based  planorm  for   Smartphones,  Tablets,  In-­‐ vehicle  Infotainment   Devices,  Smart  TVs   •  Part  of  the  Linux  Founda*on   •  Devices  from  Samsung  “later   in  2013”  
• 35.
  Ubuntu  for  Smartphones   •  Built  on  Ubuntu  Linux,   leverages  Android  Tech   •  HTML5  and  Na*ve   Apps  are  equal  ci*zens   •  Same  OS  across   devices,  same  OS   across  same  device  in   different  form  factors   •  OEM  Customizable   •  Phones  in  4Q2013/ early  2014  
• 36.
  If  you’re  confused  …  
• 37.
  Source:  hups://github.com/kumadasu/*zen-­‐history/blob/master/*zen-­‐history.pdf  
• 38.
  Not  to  forget  …  
• 39.
  Firefox  Mobile   • “Boot  to  Gecko”  -­‐>  Linux  +   Mozilla  Gecko   •  Web  as  the  Applica*on   Planorm   •  Leverages  HTML5  and   other  Web  Standards   •  Devices  launching  “soon”  
• 40.
  Sources   Google  and  Wolfram  Alpha   hup://www.embedded-­‐linux.co.uk/embedded-­‐history   hup://www.arm.linux.org.uk/docs/history.php   hup://lwn.net/Ar*cles/437162/   hup://www.linuxjournal.com/ar*cle/5755   hup://en.wikipedia.org/wiki/Carl_von_Clausewitz   hups://rt.wiki.kernel.org/index.php/Main_Page   hup://elinux.org/images/b/ba/Elc2013_Rostedt.pdf   hups://lwn.net/Ar*cles/518988/   hup://en.wikipedia.org/wiki/Sharp_Zaurus   hup://en.wikipedia.org/wiki/Nokia_770_Internet_Tablet   hups://wiki.linaro.org/WorkingGroups/ PowerManagement   hups://www.kernel.org/doc/Documenta*on/cpu-­‐ hotplug.txt   hup://elinux.org/Power_Management   hup://www.itworld.com/mobile-­‐wireless/175829/arm-­‐ and-­‐linux-­‐major-­‐construc*on-­‐ahead   hup://arstechnica.com/business/2012/08/from-­‐altair-­‐to-­‐ ipad-­‐35-­‐years-­‐of-­‐personal-­‐computer-­‐market-­‐share/   hup://www.pcworld.com/ar*cle/2042267/amd-­‐slates-­‐ first-­‐arm-­‐server-­‐chip-­‐seaule-­‐for-­‐2014.html   hup://www.slideshare.net/opersys/is-­‐android-­‐the-­‐new-­‐ embedded-­‐embedded-­‐linux-­‐at-­‐embedded-­‐wordl-­‐2013   hups://wiki.linaro.org/WorkingGroups/ PowerManagement/Doc/Architecture   hup://www.forbes.com/sites/quora/2013/05/13/what-­‐ are-­‐the-­‐major-­‐changes-­‐that-­‐android-­‐made-­‐to-­‐the-­‐linux-­‐ kernel/   hup://training.linuxfounda*on.org/free-­‐linux-­‐training/ download-­‐training-­‐materials/history-­‐of-­‐embedded-­‐linux-­‐ and-­‐best-­‐prac*ces-­‐for-­‐gemng-­‐started   hup://en.wikipedia.org/wiki/Linux_Router_Project   hup://www.extremetech.com/extreme/155392-­‐ interna*onal-­‐space-­‐sta*on-­‐switches-­‐from-­‐windows-­‐to-­‐ linux-­‐for-­‐improved-­‐reliability   hup://www.linaro.org   hup://yoctoproject.org   hup://*zen.org   hup://www.ubuntu.com/phone   hups://github.com/kumadasu/*zen-­‐history/blob/ master/*zen-­‐history.pdf   hup://www.mozilla.org/en-­‐US/firefox/os/