ARTiS, an asymmetric real-time scheduler

diff -urpN -X /export/src/patches/dontdiff -X /export/src/patches/dontdiff2 2.6.11-pfm/Documentation/artis.txt 2.6.11-artis-cvs/Documentation/artis.txt
@@ -0,0 +1,243 @@
+ARTiS - an Asymmetric Real-Time Scheduling
+http://www.lifl.fr/west/artis/
+https://gna.org/projects/artis/
+
+Eric Piel <[email protected]>
+Christophe Osuna <[email protected]>
+Julien Soula <[email protected]>
+Philippe Marquet <[email protected]>
+
+Université des Sciences et Technologies de Lille
+Laboratoire d'Informatique Fondamentale de Lille
+Last update: May 02, 2005
+------------------------------------------------------------------------------
+
+
+Overview
+--------
+
+ARTiS is a patch to the Linux scheduler directed to real-time SMP
+systems. By migrating tasks from real-time CPUs to non-real-time ones
+it can achieve very low latencies. On an 4x IA-64 we target user
+latencies always under 300µs.
+
+Detail descriptions of the implementation and measurements are available as
+research articles at: http://www.lifl.fr/west/artis .
+
+Installation
+------------
+
+ARTiS is provided as a set of Linux kernel patch. They apply against
+the vanilla Linux Kernel, available at http://www.kernel.org . Some
+versions might need to be applied against the Bull version of the
+Linux kernel. Use the command:
+ % patch -p1 < ../the-name-of-the.patch 
+
+ARTiS was designed to run on IA-64 and x86 SMP systems. You can
+compile it into the kernel by selecting:
+
+ [X] File systems -> Pseudo filesystems -> /proc file system support
+ [X] Processor types and features -> Preemptible Kernel
+ [X] Processor types and features ->   Compile the kernel with ARTiS
+ support
+
+Optionally you can select:
+
+ [X] Processor types and features ->     Compile the kernel with ARTiS
+ debugging support
+ [X] Processor types and features ->     Compile the kernel with ARTiS
+ accounting support
+
+You also need to disable the energy saving functions that could put the CPUs
+into a state with a long transition latency. For example, on IA-64 the
+pal_halt optimisation must be disabled, this can be done by appending "nohalt"
+to the kernel command line (in elilo.conf).
+
+System administration
+---------------------
+
+Before you can benefit from the ARTiS enhancements, you will have to
+boot on a Linux kernel including the patches on a SMP computer.
+
+You need to choose an appropriate CPU for the real-time program and
+set affinities accordingly via the /proc file system.
+
+The ARTiS interface provides three files:
+/proc/artis/active : contains 0 or 1 corresponding to the deactivation
+  or activation of ARTiS. It is writable.
+/proc/artis/maskrt : contains a hexadecimal number corresponding to the RT
+  CPU mask. You can change it only when ARTiS is disabled.
+/proc/artis/cpustat : gives various statistics about ARTiS, it is read-only.
+
+EXAMPLE:
+
+This is an x86 system with 4 processors. We will use the RTC interrupt
+for our real-time program, which will run on the 4th processor.
+
+   % cat /proc/interrupts
+              CPU0       CPU1       CPU2       CPU3
+     0:   16320035         14         10         11    IO-APIC-edge  timer
+     1:          8          1          0          0    IO-APIC-edge  i8042
+     2:          0          0          0          0          XT-PIC  cascade
+     8:          2          0          1    2459330    IO-APIC-edge  rtc
+    19:    9930157    1643458    1603777          0   IO-APIC-level  eth0
+    20:        133          0          1          0   IO-APIC-level  aic7xxx
+    21:         29          0          1          0   IO-APIC-level  aic7xxx
+    22:     134403          0          1          0   IO-APIC-level  aic7xxx
+   NMI:          0          0          0          0
+   LOC:   16318912   16318911   16318910   16318835
+   ERR:          0
+   MIS:          0
+   
+We need to redirect all the interrupts except the RTC one to the 3
+first processors (which means 7 as CPU mask):
+
+   % for i in 1 2 19 20 21 22 ; do echo 7 > /proc/irq/$i/smp_affinity ; done
+
+Alternatively, you can use the change_all_irqs.sh script available on
+the ARTiS webpage in the affinity-utils package. Warning, on IA-64, we
+have noticed that some IRQ may not be able to change of mask. You need
+to select another CPU than the one where the IRQ are as a RT CPU.
+
+Here we want to keep IRQ 8 (RTC) for the real-time process on the 4th CPU:
+
+   % echo 8 > /proc/irq/8/smp_affinity
+
+We want the 4th processor to be set as real-time:
+
+   % echo 8 > /proc/artis/maskrt
+
+And of course we must activate ARTiS :-)
+
+   % echo 1 > /proc/artis/active
+
+The active state of ARTiS as well as the maskrt properties can also be
+set on the command line (in elilo.conf) via respectively
+"artis_active" and "artis_maskrt=".
+
+IMPORTANT NOTE (bug workaround): As of 20050502, there seems to be a
+latency problem on SMPs. This bug appears both on x86 and IA-64, even
+when the computer is purely idle and without ARTiS. The problem: there
+is always one (and only one) CPU where the latencies are much higher
+than anywhere else (10x or 100x higher). This CPU seems to be the
+same that receive all the IRQs at boot, so you can detect it by having
+a look at /proc/interrupts. Set this CPU as an NRT CPU, that's all.
+
+
+ARTiS /proc interface
+---------------------
+
+The ARTiS activation is now realized after boot-time via the /proc
+interface.
+
+The following enables ARTiS
+
+   % echo 1 > /proc/artis/active
+
+while the following disable ARTiS
+
+   % echo 0 > /proc/artis/active
+
+
+The file /proc/artis/cpustat contains one line of statistics per CPU
+(a maximum of 8 CPUs). Every line contains in turn four columns with
+the number of migrations initiated on this CPU and some timings useful
+for debugging.
+
+
+The file /proc/artis/maskrt identifies the RT CPUs.
+
+You can specify which CPUs are real-time by writing into this file
+the mask of the CPUs.
+
+For instance the following will set 4th processor as real-time:
+
+   % echo 8 > /proc/artis/maskrt
+
+The following will set 2nd, 3rd and 4th CPUs as real-time:
+
+   % echo e > /proc/artis/maskrt
+
+
+
+
+Real-Time programming with ARTiS
+--------------------------------
+
+You will need to add some instructions to a program to make it
+real-time and to bind it to a CPU. Fortunately this is not a difficult
+task and if your application is POSIX compliant, most of them are
+probably already done. A function like the next one can be used for
+this:
+
+   #include <string.h>
+   #include <sched.h>
+
+   /* Set the current process to real-time priority.
+    * Return 0 on success, -1 on failure */
+   int set_realtime_priority(void)
+   {
+       sched_param schp;
+
+       memset(&schp, 0, sizeof(schp));
+       schp.sched_priority = sched_get_priority_max(SCHED_FIFO);
+
+       return sched_setscheduler(0, SCHED_FIFO, &schp);
+   }
+
+Setting the affinity of a process might not be correctly handled by
+the glibc, so you will have to make a syscall. The following might
+help:
+
+   #include <unistd.h>
+
+   #if defined(__ia64__)
+   #define SYS_sched_setaffinity 1231
+   #define SYS_sched_getaffinity 1232
+   #elif defined(__i386__)
+   #define SYS_sched_setaffinity 241
+   #define SYS_sched_getaffinity 242
+   #else
+   #error Architecture not supported (only ia64 an i386 are)
+   #endif
+   
+   #define sched_setaffinity(arg1, arg2, arg3) \
+           syscall(SYS_sched_setaffinity, arg1, arg2, arg3)
+   #define sched_getaffinity(arg1, arg2, arg3) \
+           syscall(SYS_sched_getaffinity, arg1, arg2, arg3)
+   
+   /* Set the affinity of the current process.
+    * Return 0 on success, -1 on failure */
+   int set_affinity(unsigned long mask)
+   {
+       return sched_setaffinity(0, sizeof(mask), &mask);
+   }
+
+With respect to the example above, our program will need to make the
+following calls somewhere:
+
+   {
+       set_realtime_priority();
+       set_affinity(8);
+   }
+
+Due to the nature of ARTiS, the only constraint compared to a usual
+POSIX RT application is the order of this two function calls. The
+priority set up must always be *before* the affinity set up
+(otherwise, the affinity cannot be guaranted).
+
+Always keep in mind that if you want your task to have obtain firm
+real-time latencies, they will have to be RT0. Such task have the
+maximum RT FIFO priority and are binded to one and only one RT CPU.
+Guaranty can be insured only for one RT0 task per CPU. More RT0 per
+CPU is possible but it is your job to check that they can all be
+executed in the same time and always repect their deadlines.
+
+In case you don't want (or cannot) recompile your application to fit
+those specific requirements for ARTiS, you can still change it to RT0
+from outside. schedtool is probably what you need, cf http://freequaos.host.sk/schedtool/.
+
+Another way to do this is to use the functions provided by libartis, a
+small package available on the ARTiS webpage. If you need a more
+elaborate example, you can have a look at realfeel-pfm, also available
+at the ARTiS webpage.

diff -urpN -X /export/src/patches/dontdiff -X /export/src/patches/dontdiff2 2.6.11-pfm/Documentation/kernel-parameters.txt 2.6.11-artis-cvs/Documentation/kernel-parameters.txt
@@ -220,6 +220,12 @@ running once the system is up.
 	arcrimi=	[HW,NET] ARCnet - "RIM I" (entirely mem-mapped) cards
 			Format: <io>,<irq>,<nodeID>
 
+	artis_active    [KNL] Activates ARTiS
+			See also Documentation/artis.txt
+
+	artis_maskrt=   [KNL] Set ARTiS real-time CPUs mask
+			See also Documentation/artis.txt
+
 	ataflop=	[HW,M68k]
 
 	atarimouse=	[HW,MOUSE] Atari Mouse

diff -urpN -X /export/src/patches/dontdiff -X /export/src/patches/dontdiff2 2.6.11-pfm/fs/proc/Makefile 2.6.11-artis-cvs/fs/proc/Makefile
@@ -8,7 +8,7 @@ proc-y			:= nommu.o task_nommu.o
 proc-$(CONFIG_MMU)	:= mmu.o task_mmu.o
 
 proc-y       += inode.o root.o base.o generic.o array.o \
-		kmsg.o proc_tty.o proc_misc.o
+		kmsg.o proc_tty.o proc_misc.o artis.o
 
 proc-$(CONFIG_PROC_KCORE)	+= kcore.o
 proc-$(CONFIG_PROC_DEVICETREE)	+= proc_devtree.o

diff -urpN -X /export/src/patches/dontdiff -X /export/src/patches/dontdiff2 2.6.11-pfm/arch/i386/Kconfig 2.6.11-artis-cvs/arch/i386/Kconfig
@@ -528,6 +528,32 @@ config PREEMPT_BKL
 	  Say Y here if you are building a kernel for a desktop system.
 	  Say N if you are unsure.
 
+config ARTIS
+	bool "Compile the kernel with ARTiS support (EXPERIMENTAL)"
+	depends on PREEMPT && SMP
+	help
+		ARTiS is an "Asymmetric Real-Time Scheduling". When activated,
+		the real-time tasks will be insured of low latencies. It works
+		by automatically migrating the normal tasks out from CPUs running
+		real-time tasks when they would trigger a preemption disable or a
+		IRQ disable. Obviously, a SMP system is required (SMT works too).
+		Note that it will not be activated by default, you need to select
+		it at boot time. For more information, see Documentation/artis.txt.
+
+config ARTIS_DEBUG
+	bool "Compile the kernel with ARTiS debugging support (EXPERIMENTAL)"
+	depends on ARTIS
+	help
+		Activate debuging code in ARTiS, you probably don't want this, excepted
+		if you want to hack on ARTiS.
+
+config ARTIS_STAT
+	bool "Compile the kernel with ARTiS accounting support (EXPERIMENTAL)"
+	depends on ARTIS
+	help
+		Activate statistics about ARTiS, available in /proc. There is no (or
+		very little) performance penalty, so you can safely say "yes" here.
+
 config X86_UP_APIC
 	bool "Local APIC support on uniprocessors" if !SMP
 	depends on !(X86_VISWS || X86_VOYAGER)

diff -urpN -X /export/src/patches/dontdiff -X /export/src/patches/dontdiff2 2.6.11-pfm/arch/i386/kernel/entry.S 2.6.11-artis-cvs/arch/i386/kernel/entry.S
@@ -361,13 +361,25 @@ common_interrupt:
 	call do_IRQ
 	jmp ret_from_intr
 
+#ifdef CONFIG_ARTIS
 #define BUILD_INTERRUPT(name, nr)	\
 ENTRY(name)				\
 	pushl $nr-256;			\
 	SAVE_ALL			\
+	call function_artis_migration_disable; \
 	movl %esp,%eax;			\
 	call smp_/**/name;		\
+	call function_artis_migration_enable; \
 	jmp ret_from_intr;
+#else
+#define BUILD_INTERRUPT(name, nr)	\
+ENTRY(name)				\
+	pushl $nr-256;			\
+	SAVE_ALL			\
+	movl %esp,%eax;			\
+	call smp_/**/name;	\
+	jmp ret_from_intr;
+#endif
 
 /* The include is where all of the SMP etc. interrupts come from */
 #include "entry_arch.h"

diff -urpN -X /export/src/patches/dontdiff -X /export/src/patches/dontdiff2 2.6.11-pfm/arch/ia64/Kconfig 2.6.11-artis-cvs/arch/ia64/Kconfig
@@ -46,6 +46,32 @@ config GENERIC_IOMAP
 	bool
 	default y
 
+config ARTIS
+	bool "Compile the kernel with ARTiS support (EXPERIMENTAL)"
+	depends on PREEMPT && SMP
+	help
+		ARTiS is an "Asymmetric Real-Time Scheduling". When activated,
+		the real-time tasks will be insured of low latencies. It works
+		by automatically migrating the normal tasks out from CPUs running
+		real-time tasks when they would trigger a preemption disable or an
+		IRQ disable. Obviously, a SMP system is required (SMT works too).
+		Note that it will not be activated by default, you need to select
+		it at boot time. For more information, see Documentation/artis.txt.
+
+config ARTIS_DEBUG
+       bool "Compile the kernel with ARTiS debugging support (EXPERIMENTAL)"
+       depends on ARTIS
+       help
+		Activate debuging code in ARTiS, you probably don't want this, excepted
+		if you want to hack on ARTiS.
+
+config ARTIS_STAT
+       bool "Compile the kernel with ARTiS accounting support (EXPERIMENTAL)"
+       depends on ARTIS
+       help
+		Activate statistics about ARTiS, available in /proc. There is no (or
+		very little) performance penalty, so you can safely say "yes" here.
+
 choice
 	prompt "System type"
 	default IA64_GENERIC